A photovoltaic power station operation and maintenance intelligent decision method and system based on multi-source data

Abstract

The application relates to the photovoltaic power station operation and maintenance technical field and discloses a photovoltaic power station operation and maintenance intelligent decision method and system based on multi-source data, which comprises the following steps: constructing an original state vector based on acquired electrical data and environmental data, and obtaining a standardized state vector through standardization processing; performing feature extraction on the standardized state vector to obtain current low-layer detailed features and current high-layer semantic features; determining an instant risk coefficient based on the current high-layer semantic features; calculating the historical similarity between the current low-layer detailed features and a historical low-layer feature vector, splicing the maximum similarity and the standardized state vector to obtain an enhanced state vector; generating a safety feature vector based on the enhanced state vector, the instant risk coefficient and compressed low-layer detailed features obtained through projection compression; generating safety action distribution parameters based on the safety feature vector; generating a control action based on the safety action distribution parameters; generating a first control instruction based on the control action and sending the first control instruction to an executing mechanism for execution.

Description

Technical Field

[0001] This application relates to the field of photovoltaic power plant operation and maintenance technology, and includes, but is not limited to, a photovoltaic power plant operation and maintenance intelligent decision-making method and system based on multi-source data. Background Technology

[0002] As the proportion of photovoltaic power generation in the energy structure continues to increase, the large-scale and centralized operation and maintenance of photovoltaic power plants faces severe challenges. Traditional operation and maintenance models rely on regular manual inspections and post-fault handling, which suffer from low efficiency, high costs, and delayed response. In recent years, intelligent algorithms, represented by deep reinforcement learning (DRL), have been introduced into the power optimization and operation control of photovoltaic power plants, aiming to achieve automated and intelligent operation and maintenance decisions.

[0003] However, in related technologies, the handling of photovoltaic power plant equipment failures is often carried out after the failure occurs, lacking foresight. Summary of the Invention

[0004] In view of this, embodiments of this application provide a method and system for intelligent decision-making in the operation and maintenance of photovoltaic power plants based on multi-source data, which at least solves the problem of delayed fault response.

[0005] The technical solution of this application embodiment is implemented as follows:

[0006] In a first aspect, embodiments of this application provide a photovoltaic power plant operation and maintenance intelligent decision-making method based on multi-source data, applied to a photovoltaic power plant operation and maintenance intelligent decision-making system based on multi-source data. The system includes a deep reinforcement learning agent, and a data acquisition module, a data preprocessing module, and an instruction execution module communicatively connected to the deep reinforcement learning agent. The deep reinforcement learning agent includes a value network, a risk gating network, a precursor memory, a state enhancement module, a security network, and a policy generation module. The method includes:

[0007] Electrical and environmental data are acquired using the data acquisition module. Based on the electrical and environmental data, an original state vector is constructed. The original state vector is then standardized using the data preprocessing module to obtain a standardized state vector.

[0008] The value network is used to extract features from the standardized state vector to obtain the current low-level detailed features and the current high-level semantic features; the risk gating network is used to determine the instantaneous risk coefficient based on the current high-level semantic features.

[0009] Using the state enhancement module, the historical similarity between the current low-level detail feature and each historical low-level feature vector in the precursor memory is calculated. The maximum similarity in the historical similarity is concatenated with the standardized state vector to obtain the enhanced state vector. The historical low-level feature vector is the low-level feature vector of the time window of the precursor of the historical fault.

[0010] Using the security network, a security feature vector is generated based on the enhanced state vector, the instantaneous risk coefficient, and the compressed low-level detail features obtained by projecting and compressing the current low-level detail features; based on the security feature vector, a security action distribution parameter containing the action mean and action variance is generated.

[0011] Using the strategy generation module, control actions are generated based on the safety action distribution parameters; using the instruction execution module, a first control instruction is generated based on the control actions, and the first control instruction is sent to the execution mechanism for execution.

[0012] Secondly, embodiments of this application provide a photovoltaic power plant operation and maintenance intelligent decision-making system based on multi-source data. The system includes a deep reinforcement learning agent, and a data acquisition module, a data preprocessing module, and an instruction execution module communicatively connected to the deep reinforcement learning agent. The deep reinforcement learning agent includes a value network, a risk gating network, a precursor memory, a state enhancement module, a security network, and a policy generation module, wherein:

[0013] The data acquisition module is used to acquire electrical data and environmental data, and construct an original state vector based on the electrical data and environmental data; the data preprocessing module is used to standardize the original state vector to obtain a standardized state vector.

[0014] The value network is used to extract features from the standardized state vector to obtain current low-level detailed features and current high-level semantic features; the risk gating network is used to determine the instantaneous risk coefficient based on the current high-level semantic features.

[0015] The state enhancement module is used to calculate the historical similarity between the current low-level detail features and each historical low-level feature vector in the precursor memory, and to concatenate the maximum similarity in the historical similarity with the standardized state vector to obtain the enhanced state vector. The historical low-level feature vector is the low-level feature vector of the time window of the precursor of the historical fault.

[0016] The security network is used to generate a security feature vector based on the enhanced state vector, the instantaneous risk coefficient, and the compressed low-level detail features obtained by projecting and compressing the current low-level detail features; and to generate a security action distribution parameter containing the action mean and action variance based on the security feature vector.

[0017] The strategy generation module is used to generate control actions based on the safety action distribution parameters; the instruction execution module is used to generate a first control instruction based on the control action and send the first control instruction to the execution mechanism for execution.

[0018] The beneficial effects of the technical solutions provided in this application include at least the following:

[0019] The system achieves automatic collection and standardized processing of operation and maintenance data through data acquisition and preprocessing modules, replacing the traditional inspection mode. Based on the synergistic effect of the value network, risk gating network, precursor memory, state enhancement module, security network, and strategy generation module in the deep strength learning agent, the system accurately extracts low-level detailed features and high-level semantic features through the value network. The state enhancement module calculates the similarity between the current low-level detailed features and the low-level feature vectors before historical faults, and concatenates the similarity with the standardized state vector to obtain the enhanced state vector. The precursor memory and state enhancement module enable accurate identification of fault precursors. Combined with the real-time risk coefficient output by the risk gating network, the security network generates reasonable safety action distribution parameters and control actions, upgrading the operation and maintenance mode from passive fault handling to proactive precursor notification and prevention. This effectively avoids power generation losses and equipment damage caused by sudden faults and ensures the stability of the power grid operation. It achieves automated and intelligent operation and maintenance decision-making, balancing the operational efficiency and safety of photovoltaic power plants, improving the accuracy and reliability of operation and maintenance decisions, and enhancing fault response efficiency. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:

[0021] Figure 1 A flowchart illustrating an intelligent decision-making method for photovoltaic power plant operation and maintenance based on multi-source data, provided in an embodiment of this application;

[0022] Figure 2 A schematic diagram illustrating the composition of a photovoltaic power plant operation and maintenance intelligent decision-making system based on multi-source data, provided in an embodiment of this application;

[0023] Figure 3 This is a schematic diagram of the composition structure of another photovoltaic power plant operation and maintenance intelligent decision-making system based on multi-source data, provided in an embodiment of this application. Detailed Implementation

[0024] 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. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0026] It should be noted that the terms "first, second, and third" used in the embodiments of this application are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0027] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0028] This application provides a method for intelligent decision-making in the operation and maintenance of photovoltaic power plants based on multi-source data, applied to electronic devices. These electronic devices include, but are not limited to, mobile phones, laptops, tablets, handheld internet devices, multimedia devices, streaming media devices, mobile internet devices, wearable devices, or other types of electronic devices. The functions implemented by this method can be achieved by a processor in the electronic device calling program code. The program code can be stored in a computer storage medium; therefore, the electronic device includes at least a processor and a storage medium. The processor can be used to process the intelligent decision-making process for the operation and maintenance of photovoltaic power plants based on multi-source data, and the memory can be used to store the data required and generated during the intelligent decision-making process.

[0029] Figure 1 This application provides a flowchart illustrating an intelligent decision-making method for photovoltaic power plant operation and maintenance based on multi-source data, applicable to, for example... Figure 2 The photovoltaic power plant operation and maintenance intelligent decision-making system shown is based on multi-source data. The system includes a deep reinforcement learning agent, and a data acquisition module, a data preprocessing module, and an instruction execution module communicatively connected to the deep reinforcement learning agent. The deep reinforcement learning agent includes a value network, a risk gating network, a precursor memory, a state enhancement module, a security network, and a policy generation module, as shown below. Figure 1 As shown, the method includes at least the following steps:

[0030] Step S110: Use the data acquisition module to acquire electrical data and environmental data, and construct an original state vector based on the electrical data and environmental data; use the data preprocessing module to standardize the original state vector to obtain a standardized state vector.

[0031] The electrical data refers to the core operating electrical data of the photovoltaic power station, reflecting the grid-connected output characteristics, power quality, and operational safety status of the photovoltaic power station; the environmental data refers to the core environmental variables that affect the power generation efficiency of the photovoltaic modules and the scattering status of the equipment in the photovoltaic power station; the standardization process is used to eliminate the influence of different physical quantities and to accelerate the training convergence speed and improve the training stability during the training process of the neural network.

[0032] Step S120: Use the value network to extract features from the standardized state vector to obtain the current low-level detailed features and the current high-level semantic features; use the risk gating network to determine the instantaneous risk coefficient based on the current high-level semantic features.

[0033] The value network is used to perform hierarchical feature mining and transformation on the standardized state vector, and finally outputs two types of features with different levels of abstraction and different downstream uses, namely the current low-level detailed features and the current high-level semantic features. The risk gating network can perform risk quantification based on the current high-level semantic features output by the value network to obtain an instantaneous risk coefficient.

[0034] Step S130: Using the state enhancement module, calculate the historical similarity between the current low-level detail feature and each historical low-level feature vector in the precursor memory bank. Concatenate the maximum similarity in the historical similarity with the standardized state vector to obtain the enhanced state vector. The historical low-level feature vector is the low-level feature vector of the time window of the precursor of the historical fault.

[0035] The precursor memory stores a set of historical low-level feature vectors within specific time windows preceding confirmed fault events. The historical similarity is used to quantify the similarity between the current operating state and historical precursor states. A higher historical similarity indicates that the current state is closer to the pattern preceding a historical fault, meaning that the risk inferred based on historical experience is higher. This precursor memory is stored in non-volatile memory, and only query operations are performed during the online inference phase; no writing or updating is performed. The state enhancement module integrates historical experience risk information into the decision state by concatenating the maximum similarity with the standardized state vector, thereby achieving state enhancement.

[0036] Step S140: Using the security network, based on the enhanced state vector, the instantaneous risk coefficient, and the compressed low-level detail features obtained by projecting and compressing the current low-level detail features, a security feature vector is generated; based on the security feature vector, a security action distribution parameter containing the action mean and action variance is generated.

[0037] The enhanced state vector is used to provide correlation information of historical fault precursors, the instantaneous risk coefficient is used to provide a quantitative indicator of the current risk, the compressed low-level detail features are used to provide fine-grained physical details of the current operation of the photovoltaic power station, and the safety network is used to integrate the enhanced state vector, the instantaneous risk coefficient, and the compressed low-level detail features into a safety feature vector, and then transform it into safety action distribution parameters for subsequent generation of conservative and safe photovoltaic power station adjustment actions.

[0038] Step S150: Using the strategy generation module, a control action is generated based on the safety action distribution parameters; using the instruction execution module, a first control instruction is generated based on the control action, and the first control instruction is sent to the execution mechanism for execution.

[0039] The strategy generation module can convert the safety action distribution parameters of the security network into specific control actions (i.e., photovoltaic power station adjustment actions), such as power reduction and voltage adjustment. The instruction execution module converts the control actions into first control instructions for the photovoltaic power station's inverters, combiner boxes, and other actuators. The actuators complete the actual adjustment operations according to the first control instructions, ultimately achieving conservative preventive control of the photovoltaic power station under risk conditions.

[0040] In the above embodiments, the automatic collection and standardized processing of operation and maintenance data are achieved through the data acquisition module and the data preprocessing module, replacing the traditional inspection mode. Based on the synergistic effect of the value network, risk gating network, precursor memory, state enhancement module, security network and strategy generation module in the deep strength learning agent, the value network accurately extracts low-level detailed features and high-level semantic features. The state enhancement module calculates the similarity between the current low-level detailed features and the low-level feature vectors before the historical faults, and concatenates the similarity with the standardized state vector to obtain the enhanced state vector. The precursor memory and the state enhancement module achieve accurate identification of fault precursors. Combined with the real-time risk coefficient output by the risk gating network, the security network generates reasonable safety action distribution parameters and control actions, upgrading the operation and maintenance mode from passive fault handling to proactive precursor notification and prevention. This effectively avoids power generation losses and equipment damage caused by sudden faults and ensures the stability of grid operation. It realizes the automation and intelligence of operation and maintenance decision-making, takes into account the operating efficiency and safety of photovoltaic power plants, improves the accuracy and reliability of operation and maintenance decisions, and improves fault response efficiency.

[0041] In some embodiments, the electrical data includes the total active power, total reactive power, and grid connection point voltage output from the AC side of the photovoltaic power station inverter; the environmental data includes ambient temperature and light intensity; and step S110, "acquiring electrical and environmental data using the data acquisition module and constructing an original state vector based on the electrical and environmental data," includes:

[0042] Step S1101: Based on the total active power, the total reactive power, the grid connection point voltage, the ambient temperature, and the light intensity, construct a five-dimensional original state vector;

[0043] The data acquisition module is used to acquire and standardize the core operating electrical data and environmental status data of the photovoltaic power station. First, the system synchronously acquires multi-source raw data from the monitoring and data acquisition system of the photovoltaic power station and the meteorological station at a fixed frequency (e.g., 1Hz) to construct the raw state vector containing five key dimensions as shown in the following formula (1). :

[0044] Formula (1);

[0045] in, and They represent in The total active power and total reactive power output of the inverter on the AC side of the photovoltaic power station at any time directly reflect the real-time energy scale and power factor of the power station to the grid. for The voltage at the grid connection point at any given time is a key indicator for measuring the power quality and operational safety of the grid connection point. for The ambient temperature at any given time for The intensity of sunlight at any given time is a key environmental variable affecting the power generation efficiency of photovoltaic modules and the heat dissipation status of the equipment.

[0046] Step S110, "using the data preprocessing module to standardize the original state vector to obtain a standardized state vector," includes:

[0047] Step S1102: Using the data preprocessing module, based on the first mean vector and the first standard deviation vector obtained from the statistical characteristics of the offline training dataset, linear transformation is performed on the data of each dimension of the original state vector to obtain the standardized state vector.

[0048] Among them, the original state vector can be Standardization is performed dimensionally to eliminate the influence of dimensions and accelerate convergence. Specifically, the first mean vector is calculated independently for each dimension using the statistical properties of the offline training dataset. and the first standard deviation vector The original state vector at each moment is linearly transformed dimension by dimension to form the standardized vector shown in formula (2). :

[0049] Formula (2);

[0050] This operation can convert the data in each dimension into a distribution with a mean of 0 and a standard deviation of 1.

[0051] In the above embodiments, by selecting core electrical data such as total active power, total reactive power, and grid connection point voltage, as well as key environmental data such as ambient temperature and light intensity, a five-dimensional original state vector is constructed, ensuring that the original state vector can comprehensively and accurately represent the operating status of the power station. At the same time, standardization processing is performed based on the first mean vector and the first standard deviation vector obtained from the offline training dataset, which improves the consistency between real-time data and training data, eliminates the interference caused by differences in the dimensions of data from different dimensions, and provides an accurate and reliable data foundation for subsequent value network feature extraction, risk assessment, and decision generation, thereby improving the stability and decision accuracy of the entire intelligent decision-making system.

[0052] In some embodiments, the value network includes a shared feature layer and a high-level feature layer. Step S120, "using the value network to extract features from the standardized state vector to obtain current low-level detailed features and current high-level semantic features," includes:

[0053] Step S1201: The normalized state vector is subjected to initial feature transformation and fusion using the fully connected layer of the shared feature layer, and the feature is activated using the ReLU activation function of the shared feature layer to obtain the current low-level detailed features;

[0054] Among them, the standardized vectors output by the data preprocessing module Value Network For deep feature mining, the value network adopts a dual-branch architecture with a multilayer perceptron. The activation values ​​of different levels within the network are designed as feature extractors, which are used to capture operational mode information at different levels of abstraction.

[0055] In the shared feature layer, input First, initial feature transformation and fusion are performed using a fully connected layer with 128 neurons. Then, the output of this fully connected layer is activated and nonlinearly mapped using the ReLU activation function to generate low-level detailed features. This low-level detail feature encodes the positive activation pattern after initial fusion of the original states, preserving the detailed information after linear combination of the original states. To ensure the stability of the low-level detail feature, a gradient stopping mechanism is applied to this feature branch during the training of the main task of the value network. The parameters are treated as constants when updating the value network parameters.

[0056] Step S1202: The current low-level detail features are processed sequentially through the first fully connected layer, the ReLU activation function, the second fully connected layer, and the ReLU activation function of the high-level feature layer. Through layer-by-layer nonlinear transformation and feature activation, the current high-level semantic features are obtained.

[0057] In the high-level feature layer, the output of the shared feature layer passes sequentially through two fully connected layers, each containing 256 neurons. Each fully connected layer is followed by a ReLU activation function for non-linear activation, achieving a higher-order non-linear transformation of feature abstraction and learning the complex patterns inherent in the learning state. Finally, the output of the last fully connected layer generates high-level semantic features. This high-level semantic feature encodes deeper semantic information about the system's operational health and potential risks.

[0058] It should be noted that the value network also includes a value output layer, which is used to map high-level semantic features into scalar-form state value estimates during the model training phase of the deep reinforcement learning agent. This is used to provide a basis for evaluating the state value of the agent's strategy optimization.

[0059] In the above embodiments, the initial transformation and activation of the original standardized state vector are achieved through a fully connected layer with a shared feature layer and a ReLU activation function, accurately extracting low-level detailed features that reflect the subtle operating state of the power plant. Then, through layer-by-layer nonlinear transformation and activation of the high-level feature layer, high-level semantic features that reflect the overall operating trend of the power plant are further extracted, achieving accurate separation and extraction of features at different levels. The low-level detailed features provide support for fault precursor identification and historical similarity calculation, while the high-level semantic features provide a basis for real-time risk coefficient calculation. The two work together to improve the accuracy of risk perception, providing high-quality feature support for the generation of subsequent safety strategies and basic strategies, and further ensuring the reliability of intelligent decision-making.

[0060] In some embodiments, the risk gating network To quantify the operational status of photovoltaic power plants, this network is a lightweight, three-layer fully connected sensor. Its computational flow is as follows: The first layer performs a linear transformation. The weight matrix Bias vector The transformed output Features obtained by ReLU activation function This process completes the initial feature dimensionality reduction and nonlinear mapping; the second layer further compresses the feature dimensions and performs a linear transformation. The weight matrix Bias vector And then process it again using the ReLU activation function to obtain This layer learns higher-order abstract patterns related to risk from high-level semantic features; finally, the output layer performs a linear transformation. The weight matrix Bias vector Result in scalar output Apply the Sigmoid activation function afterwards The output is constrained to the interval [0, 1] to obtain the instantaneous risk coefficient. .

[0061] Regarding the training mechanism, the instantaneous risk coefficient... Without independent labels, its training is achieved through an end-to-end reinforcement learning framework using deep reinforcement learning agents. The parameter optimization of the risk gating network is achieved by providing gradient signals through the η4 and η5 terms in the reward function, and the end-to-end training drives the network to learn the mapping from the operating state of the photovoltaic power plant to the risk probability.

[0062] The entire calculation process can be summarized as shown in the following formula (3):

[0063] ;

[0064] This instant risk factor It represents the probability of an immediate occurrence of a risk determined by the agent based on the high-level semantic features of the current state. The closer the value is to 1, the more likely the agent's judgment system is in a high-risk operating range.

[0065] In some embodiments, the method further includes:

[0066] Step S121: Obtain all confirmed actual fault events during the operation of the system;

[0067] Step S122: Mark the time point of each actual fault event;

[0068] Step S123: Based on the fault precursor time window, extract the historical low-level feature vector within the corresponding fault precursor time window before the fault occurrence time point;

[0069] Step S124: Store each of the historical low-level feature vectors in the precursor memory bank in an orderly manner according to the timestamp.

[0070] Low-level detail features based on value network output This precursor memory stores a set of low-level feature vectors within a specific time window (e.g., 30 to 120 seconds before a confirmed failure) preceding historically confirmed failure events. ,Right now During training, at each iteration, the system calculates the current low-level detailed features. The cosine similarity with each historical low-level feature vector in the precursor memory bank.

[0071] Then, take the maximum similarity among all historical similarities. If the precursor memory is empty (e.g., in the initial training phase), then proceed with subsequent splicing. This value quantifies the similarity between the current operating state and historical pre-failure states. The closer it is to 1, the closer the current state is to the pattern before a certain failure occurred in the past, meaning the higher the risk inferred from historical experience.

[0072] The precursor memory is stored in non-volatile memory. During the online inference phase (i.e., the application phase), only query operations are performed, without writing or updating.

[0073] Based on the maximum similarity in historical similarity and standardized state vector By splicing together historical experience and risk information, the decision state is enhanced by integrating it into the decision state. The enhanced state vector containing historical risk information is constructed by connecting the basic sequence according to the dimensions, as shown in the following formula (4). :

[0074] Formula (4);

[0075] In the above embodiments, by acquiring confirmed actual fault events and extracting historical low-level feature vectors within the time window of the precursors before the fault occurs, it is ensured that the samples in the memory are all real and valid precursor features of faults, eliminating the interference of invalid samples. At the same time, the samples are stored in an orderly manner according to timestamps, and the memory is constructed using samples verified by faults, which greatly improves the accuracy of historical similarity calculation. This enables the state enhancement module to accurately identify whether there are precursors of faults in the current power plant operation state, further strengthening the system's predictability of potential faults, providing a reliable historical reference for the generation of safety strategies, helping to achieve proactive fault prevention and control, and reducing losses caused by sudden faults.

[0076] In some embodiments, such as Figure 3 As shown, the deep reinforcement learning agent also includes a base network, and the method further includes:

[0077] Step S131: Using the base network, generate basic action distribution parameters based on the enhanced state vector;

[0078] Based on enhanced state vector The basic network is designed to maximize power generation efficiency while ensuring the basic operational safety of the photovoltaic power plant. This network is a three-layer fully connected neural network, and its computational flow is as follows: The first layer receives the augmentation state vector. As input, it undergoes feature transformation to generate an initial feature vector after preliminary mapping. The second layer expands the feature dimension to 256 dimensions to achieve higher-order feature representation, resulting in an intermediate feature vector. Finally, the output layer performs a linear transformation. Among them, the weight matrix Mapping 256-dimensional features into a 2-dimensional output vector. ,in, For the output vector The first dimension element, For the output vector The second dimension of the output vector is used to decompose the vector, and the first element is the action mean. The second element is the standard deviation of the logarithm of the actions. .

[0079] At this point, the basic strategy outputs basic action distribution parameters oriented towards efficiency. The Gaussian policy distribution corresponding to the basic action is defined by this parameter. .

[0080] Step S150, "using the strategy generation module to generate control actions based on the safety action distribution parameters," includes:

[0081] Step S1501: Using the strategy generation module, a control action is generated based on the basic action distribution parameters, the safety action distribution parameters, and the instantaneous risk coefficient.

[0082] The security network and the basic network adopt a parallel policy decision-making architecture, and the security network is based on an enhanced state vector. 1-dimensional instantaneous risk coefficient and projection matrix 8D compressed low-level detail features after online real-time compression The safety action distribution parameters are generated. The goal of the safety network is to generate conservative preventive actions when a risk is perceived, as shown in formula (5) below. The safety feature vector at time t is obtained by feature concatenation. :

[0083] Formula (5);

[0084] The compressed projection aims to preserve key details in the current low-level feature set while controlling the input dimension to avoid overly complex networks. Subsequently, the secure feature vectors are... Enter secure network The three-layer fully connected structure performs a linear transformation on the first layer, mapping features from 15 to 64 dimensions. The weight matrix The second layer further maps the features to 32 dimensions and performs a linear transformation. The weight matrix The output layer performs a linear transformation. The weight matrix Mapping 32-dimensional features into 2-dimensional output vectors Similarly, the output vector is decomposed, where the first element is the mean of the safety actions. The second element is the logarithmic standard deviation of the safe actions. This leads to the standard deviation of safe actions. At this point, the security network outputs conservative security action distribution parameters. The parameter defines the conservative Gaussian distribution corresponding to the safety action. .

[0085] In the above embodiments, the basic network generates basic action distribution parameters based on enhanced state vectors with the goal of maximizing power generation efficiency. Combined with conservative safety action distribution parameters generated by the safety network, the strategy generation module determines the weights of the two strategies based on the real-time risk coefficient and integrates them. This achieves a dynamic balance between efficiency priority and safety control. In low-risk situations, the basic strategy is emphasized to ensure power generation efficiency, while in high-risk situations, the safety strategy is emphasized to avoid faults. This avoids the limitations of a single strategy, improves the flexibility and rationality of operation and maintenance decisions, further optimizes the operating efficiency of photovoltaic power plants, and makes the generation of control actions more targeted, adapting to the operation and maintenance needs under different risk scenarios.

[0086] In some embodiments, step S1501, "using the strategy generation module to generate control actions based on the basic action distribution parameters, the safety action distribution parameters, and the instantaneous risk coefficient," includes:

[0087] Step S15011: Based on the instantaneous risk coefficient, determine the first mixed weight of the basic action distribution parameter and the second mixed weight of the safe action distribution parameter;

[0088] Wherein, the first mixed weight can be The second mixing weight can be .

[0089] Step S15012: Generate a hybrid action distribution based on the basic action distribution parameters, the first hybrid weight, the safety action distribution parameters, and the second hybrid weight;

[0090] Step S15013: Sample from the mixed action distribution to obtain the control action.

[0091] Based on the basic action distribution parameters output by the basic network and the security action distribution parameters output by the security network and immediate risk coefficient The strategy generation module aims to generate the final action distribution for execution. In practice, it uses... To dynamically mix the weights, linear interpolation is performed on the action distribution parameters of the base network and the security network. To avoid the variance nondefinite problem caused by directly mixing the standard deviations, this application performs linear interpolation in the logarithmic standard deviation space as shown in formula (6):

[0092] Formula (6);

[0093] Simultaneously, as shown in formula (7) below, the motion mean parameters are linearly mixed synchronously to obtain the final mixed mean. As shown in formula (8) below, the final mixed standard deviation is obtained by performing an exponential operation on the logarithmic standard deviation. :

[0094] Formula (7);

[0095] Formula (8);

[0096] The hybrid formulas (7) and (8) give the system the ability to flexibly adjust strategy preferences according to real-time risk conditions, when the real-time risk coefficient... When the value approaches 0, it indicates that the system is currently in a low-risk operating state, at which point the final action distribution is almost entirely determined by the output of the basic network. The system is determined to operate with the goal of maximizing power generation efficiency; when the instantaneous risk coefficient... As the weights increase, the output weights of the security network increase accordingly, and the final action distribution gradually becomes more conservative. Offset; when When the value approaches 1, it indicates that the system is in an extremely high-risk operating state. Ultimately, the distribution of actions is almost entirely dominated by the security network to ensure that maximum preventative measures (such as significant de-rated measures) are taken to proactively avoid potential failures.

[0097] After the strategies are mixed, the system obtains a comprehensive Gaussian action distribution (i.e., a mixed action distribution). To balance the determinism of the strategy with the necessary exploratory nature, the specific final action distribution (i.e., control action) is obtained by sampling from this distribution, as shown in the following formula (9). :

[0098] ;

[0099] In the above embodiments, by dynamically determining the weights of the basic action distribution parameters and the safety action distribution parameters based on the real-time risk coefficient, the dynamic and rational weight allocation is achieved, ensuring accurate matching between risk states and strategy weights. Then, a hybrid action distribution is generated through linear interpolation, and control actions are sampled from the distribution. This balances the determinism of strategy execution with the necessity for the agent to explore the environment, avoiding the rigidity or excessive randomness caused by directly determining actions. The entire process is standardized and traceable, further improving the accuracy and rationality of control action generation, ensuring that the system can adopt the optimal control strategy under different risk scenarios, balancing power generation efficiency and operational safety.

[0100] In some embodiments, the instruction execution module is based on the final action distribution output by the strategy generation module. The system issues power control commands to the actuators of the photovoltaic power station; after receiving and executing the commands, the environment (simulation model or real power station) feeds back the standardized state for the next moment. And a binary fault flag , This indicates that a fault has occurred. This indicates a normal state. Based on this feedback and the internal signals generated during this decision-making cycle, the system performs two key post-processing operations: calculating the immediate reward function to guide policy optimization. And dynamically update the risk precursor memory based on the fault situation.

[0101] The system calculates the immediate reward function used to drive the training of the deep reinforcement learning agent. The reward function is carefully designed to guide the agent to learn a comprehensive strategy that balances efficiency, safety, and predictability through a multi-objective weighted sum. Its specific structure is shown in formula (10), and its design aims to guide the agent's behavior according to different scenarios:

[0102] Formula (10);

[0103] in, This represents the total active power actually generated by the photovoltaic power station at the current moment. This represents the theoretical maximum power generation capacity that the power station can achieve under current environmental conditions (sunlight, temperature). The voltage at the grid connection point. This is a fault indicator. , , These are all indicator functions; they take a value of 1 when the condition is true, and 0 otherwise. This indicates the percentage reduction in the security policy of the secure network compared to the basic policy of the basic network. The power generation setting value corresponding to the action output by the security network. The power generation setting value corresponding to the action output by the backbone strategy network (i.e., the basic network); the risk-action matching function is shown in the following formula (11):

[0104] Formula (11);

[0105] This function specifies the reduction ratio recommended by the security policy. With real-time risk coefficient A certain proportion ( For example, scaling factor The maximum value is achieved when the risk level is matched. This encourages security strategies to take preventative actions that are neither excessive nor insufficient, commensurate with the current level of risk. To achieve the maximum similarity, and For threshold (e.g.) ). to These are non-negative weighting coefficients used to balance the importance of various objectives. Power generation efficiency weighting Voltage stability weighting, Fault penalty weighting Risk perception weight, Security policy weighting; risk perception reward ( Rewards are only issued when the system is not experiencing failures and its historical similarity is not high-risk, to avoid inappropriate incentives for the system due to a persistently high-risk state. (Security policy effectiveness reward) This only triggers when there is no failure and the immediate risk exceeds the threshold. The policy generation module outputs... This is the power setting value, which, after decoding, generates the actual power command P. cmd = ×P max , where P max Calculated using a standard PV model based on current irradiance and ambient temperature. and The average values ​​of the outputs of the security policy and the basic policy, respectively. It was obtained through the same decoding process.

[0106] The significance of this reward function lies in the following: firstly, it encourages power plants to maintain high power generation efficiency; secondly, it penalizes voltage deviations from rated values ​​to ensure grid stability; thirdly, it imposes severe penalties for fault occurrences, forcing the agent to avoid dangerous operations; and fourthly, it rewards successful risk perception (i.e., higher output) when no fault occurs. The fifth item is a newly added core reward item, which provides positive incentives to encourage increased risk sensitivity. At that time, if the power recommended by the security policy Power lower than recommended by the baseline strategy Then, based on the reduction ratio and risk coefficient Providing additional incentives directly encourages security strategies to take effective and significant preventative depreciation actions in high-risk scenarios.

[0107] At the same time, the system executes the precursor memory update mechanism. Real-time updates are only triggered during the training phase and when a fault occurs. Real-time updates are not performed during application, but the contents of the precursor memory are updated after a fixed period of time at the application interval.

[0108] The precursor memory update process includes: when a fault is triggered, reverse-engineering the continuous normal period of at least 30 seconds prior to the fault; uniformly sampling 10 low-level detail features from the last 30 seconds of this period and storing them in the precursor memory; the precursor memory adopts a first-in, first-out (FIFO) management principle, with a capacity N∈[100,500]. If the total number exceeds [a certain threshold] after adding new samples... If the earliest stored sample is not found, it will be automatically removed. This mechanism ensures that the precursor memory only records verified feature patterns from when the system was in continuous normal operation before the failure, or the state during or after the failure. This guarantees that subsequent online applications can be based on historical similarity. Calculate the reliability and value of risk warning.

[0109] After completing the decision execution, environmental feedback reception, and immediate reward calculation for the current moment, the system employs an experience replay mechanism to sample training data in batches from the experience replay pool. It then uses the standard PPO (Proximal Policy Optimization) algorithm to synchronously update the network parameters, first calculating the generalized dominance function estimate. Then, based on the PPO-Clip loss function, the parameters of the value network, the basic network, the security network, and the risk gating network are updated synchronously.

[0110] In this application, each network adopts a parallel architecture design, and the gradient calculation paths of each network are independent, avoiding the complex second-order derivative problem that may be caused by network concatenation, thereby ensuring the clarity of gradient propagation and the stability of the training process. The entire update process is repeated after each sampling batch until the model converges or reaches the preset number of training rounds, ultimately enabling the agent to maximize power generation efficiency while ensuring safety, and possessing the endogenous ability to predict potential faults.

[0111] It should be noted that all network parameters are updated synchronously within the same training step, with the parameters of the four main networks being computed and updated in parallel. Specifically, the loss calculations for all four networks are based on the same batch of empirical data and the same advantage function estimation. In each parameter update iteration, the value network loss is first calculated and its parameters are updated. ,get Then, the advantage function is recalculated using the updated value network (or the old version), and the losses of the three policy networks (base network, security network, and risk-gated network) are calculated simultaneously, with their parameters updated synchronously. This synchronous update method ensures that each module collaborates and optimizes in real time during the learning process, avoiding inconsistencies in knowledge or difficulties in convergence that may occur during phased training. After the agent training is completed, the final converged network parameters are obtained. , , , .

[0112] In some embodiments, the system further includes an early warning decision module, and the method further includes:

[0113] Step S161: Using the early warning decision module, when the instantaneous risk coefficient is greater than the first risk threshold and less than or equal to the second risk threshold, determine the early warning level as a state of concern, record the current state, and output prompt information on the monitoring platform.

[0114] Among them, the system generates the final control action. At the same time, the immediate risk coefficient and maximum similarity It performs continuous monitoring and multi-level threshold judgment, and triggers corresponding early warning and control actions accordingly; the early warning decision module will... The system compares the data in real time with preset thresholds and generates tiered warnings based on preset logic. The first risk threshold can be 0.6, and the second risk threshold can be 0.7. If the value is less than or equal to 0.7, meaning a higher-level warning has not been triggered, the system determines the warning level to be "attention," records the current status, and sends a notification to the monitoring platform.

[0115] Step S162: Using the early warning decision module, if the instantaneous risk coefficient is greater than the second risk threshold and less than or equal to the third risk threshold, or if the maximum similarity is greater than the first similarity threshold and less than or equal to the second similarity threshold, determine the early warning level as a high-risk state, generate a first early warning information containing risk information and suggested measures, and push the first early warning information to the operation and maintenance platform or mobile application.

[0116] Wherein, the third risk threshold can be 0.85, the first similarity threshold can be 0.8, and the second similarity threshold can be 0.85, then if or If the value is less than or equal to 0.85, the system determines the warning level to be high-risk, and the system automatically generates a warning value including a timestamp and a risk value. The system provides structured early warning information, including the best-matched historical failure mode information and recommended measures (such as suggesting a reduction in the rate of operation to 80%), and pushes it to the operation and maintenance platform and mobile application.

[0117] Step S163: Using the early warning decision module, when the immediate risk coefficient is greater than the third risk threshold and the maximum similarity is greater than the second similarity threshold, determine the early warning level as an emergency state, generate a second early warning information containing risk information and suggested measures, automatically execute a second control command to reduce output power according to the preset strategy configuration, and push the second early warning information to the operation and maintenance platform or mobile application.

[0118] like and If the system determines the warning level to be an emergency, it will generate the highest-level warning information and automatically execute a throttling control command according to the policy configuration. All warning information is synchronously distributed to the integrated management and control platform, mobile operation and maintenance application, and log system through a unified interface. Simultaneously, the control execution module will output the final action based on the policy. The commands are converted into specific, device-recognizable control instructions (such as inverter power setpoint instructions based on Modbus RTU or IEC 61850 protocols) and sent to the corresponding actuators (such as inverters and combiner boxes) of the photovoltaic power station via a secure communication link to complete early warning and control. The information is then reported to the staff through an application.

[0119] In the above embodiments, by comparing the real-time risk coefficient, maximum similarity, and preset multi-level thresholds, three early warning levels—attention, high risk, and emergency—are defined. Differentiated handling measures are taken for different levels. In the attention state, only a prompt is recorded; in the high-risk state, early warning information and suggested measures are pushed; and in the emergency state, de-rated control is automatically implemented and an early warning is pushed. This achieves hierarchical, precise, and automated risk early warning. It avoids the waste of operation and maintenance resources caused by excessive early warnings, prevents the expansion of faults due to delayed early warnings, provides clear handling guidelines for operation and maintenance personnel, shortens the fault handling cycle, improves operation and maintenance response efficiency, further strengthens the fault prevention and control capabilities of photovoltaic power plants, and ensures the safe and stable operation of power plants.

[0120] Based on the foregoing embodiments, this application provides a photovoltaic power plant operation and maintenance intelligent decision-making system based on multi-source data. The system includes various modules and sub-modules, and each unit of each sub-module can be implemented by a processor in an electronic device; of course, it can also be implemented by specific logic circuits. In the implementation process, the processor can be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), etc.

[0121] Figure 2 A schematic diagram of the composition structure of a photovoltaic power plant operation and maintenance intelligent decision-making system based on multi-source data is provided in this application embodiment, as shown below. Figure 2 As shown, the system 200 includes:

[0122] A deep reinforcement learning agent, and a data acquisition module, a data preprocessing module, and an instruction execution module communicatively connected to the deep reinforcement learning agent. The deep reinforcement learning agent includes a value network, a risk gating network, a precursor memory, a state enhancement module, a security network, and a policy generation module, wherein:

[0123] The data acquisition module is used to acquire electrical data and environmental data, and construct an original state vector based on the electrical data and environmental data; the data preprocessing module is used to standardize the original state vector to obtain a standardized state vector.

[0124] The value network is used to extract features from the standardized state vector to obtain current low-level detailed features and current high-level semantic features; the risk gating network is used to determine the instantaneous risk coefficient based on the current high-level semantic features.

[0125] The state enhancement module is used to calculate the historical similarity between the current low-level detail features and each historical low-level feature vector in the precursor memory, and to concatenate the maximum similarity in the historical similarity with the standardized state vector to obtain the enhanced state vector. The historical low-level feature vector is the low-level feature vector of the time window of the precursor of the historical fault.

[0126] The security network is used to generate a security feature vector based on the enhanced state vector, the instantaneous risk coefficient, and the compressed low-level detail features obtained by projecting and compressing the current low-level detail features; and to generate a security action distribution parameter containing the action mean and action variance based on the security feature vector.

[0127] The strategy generation module is used to generate control actions based on the safety action distribution parameters; the instruction execution module is used to generate a first control instruction based on the control action and send the first control instruction to the execution mechanism for execution.

[0128] In some possible embodiments, such as Figure 3 As shown, the deep reinforcement learning agent further includes:

[0129] A base network is used to generate basic action distribution parameters based on the enhanced state vector;

[0130] The strategy generation module is used to generate control actions based on the basic action distribution parameters, the safety action distribution parameters, and the instantaneous risk coefficient.

[0131] In some possible embodiments, the system further includes an early warning decision module, which is used to determine the early warning level as a state of concern when the instantaneous risk coefficient is greater than a first risk threshold and less than or equal to a second risk threshold, record the current state, and output prompt information on the monitoring platform.

[0132] The early warning decision module is used to determine the early warning level as a high-risk state when the instantaneous risk coefficient is greater than the second risk threshold and less than or equal to the third risk threshold, or when the maximum similarity is greater than the first similarity threshold and less than or equal to the second similarity threshold, generate first early warning information containing risk information and suggested measures, and push the first early warning information to the operation and maintenance platform or mobile application.

[0133] The early warning decision module is used to determine the early warning level as an emergency state when the immediate risk coefficient is greater than the third risk threshold and the maximum similarity is greater than the second similarity threshold, generate a second early warning message containing risk information and suggested measures, automatically execute a second control command to reduce output power according to a preset strategy configuration, and push the second early warning message to the operation and maintenance platform or mobile application.

[0134] In some possible embodiments, the electrical data includes the total active power, total reactive power, and grid connection point voltage output by the AC side of the photovoltaic power station inverter; the environmental data includes ambient temperature and light intensity; and the data acquisition module is used to construct a five-dimensional original state vector based on the total active power, the total reactive power, the grid connection point voltage, the ambient temperature, and the light intensity.

[0135] The data preprocessing module is used to perform linear transformations on the data of each dimension of the original state vector based on the first mean vector and the first standard deviation vector obtained from the statistical characteristics of the offline training dataset, so as to obtain a standardized state vector.

[0136] In some possible embodiments, the value network includes a shared feature layer and a high-level feature layer. The value network is used to perform initial feature transformation and fusion on the standardized state vector using the fully connected layer of the shared feature layer, and to perform feature activation using the ReLU activation function of the shared feature layer to obtain the current low-level detailed features. The current low-level detailed features are then processed sequentially through the first fully connected layer, the ReLU activation function, the second fully connected layer, and the ReLU activation function of the high-level feature layer. Through layer-by-layer nonlinear transformation and feature activation, the current high-level semantic features are obtained.

[0137] In some possible embodiments, the precursor memory is also used to acquire all confirmed actual fault events during the operation of the system; mark the fault occurrence time point for each actual fault event; based on the fault precursor time window, extract the historical low-level feature vector within the corresponding fault precursor time window before the fault occurrence time point; and store each historical low-level feature vector in the precursor memory in an orderly manner according to the timestamp.

[0138] In some possible embodiments, the strategy generation module is configured to determine a first mixed weight of the basic action distribution parameters and a second mixed weight of the safety action distribution parameters based on the instantaneous risk coefficient; generate a mixed action distribution based on the basic action distribution parameters, the first mixed weight, the safety action distribution parameters, and the second mixed weight; and sample from the mixed action distribution to obtain control actions.

[0139] This application provides a photovoltaic power plant operation and maintenance intelligent decision-making and early warning system based on multi-source data, the system architecture of which is as follows: Figure 3 As shown, the online operation process of this system is as follows:

[0140] First, the model and knowledge base are loaded, including the final network parameters after training and convergence. , , , And the final static precursor memory bank accumulated and solidified during the training phase. The online application of the system is deployed on the edge computing node of the photovoltaic power station. The hardware architecture includes a 1Hz data acquisition module (supporting Modbus TCP / IEC 61850 protocol), an AI inference unit (using NVIDIA Jetson AGX Orin or equivalent computing power hardware), a real-time communication switch (supporting IEEE 1588 precise time synchronization), and an actuator interface.

[0141] Next, the data acquisition module collects real-time operating data of the photovoltaic power station at a fixed frequency (e.g., 1Hz). The data preprocessing module performs basic preprocessing on the collected real-time data, using the same standardized parameters as in the model training phase. After completing the data standardization process, we obtain .

[0142] Next, after obtaining the real-time standardized state Then, it is fed into the pre-trained value network. Perform forward propagation to extract high-level semantic features With low-level detail features ,Right now Subsequently, the system computes two types of risk perception signals in parallel, incorporating high-level semantic features. Input a fixed risk gating network Obtain the immediate risk coefficient At the same time, low-level detailed features With solidified precursor memory bank Cosine similarity is calculated for all historical samples, and the maximum value is taken as the maximum similarity. Based on calculations System build enhancement status It serves as input to the foundational strategy (i.e., the basic network); simultaneously, it constructs a dedicated input feature vector for the security strategy (i.e., the security network). Next, the system executes the basic policy and security policy in parallel: Input basic network Obtain the basic motion distribution parameters ;Will Enter secure network Obtain safety action distribution parameters Finally, based on the immediate risk factor... The outputs of the two strategies are linearly mixed to generate the parameters of the final action distribution. and and from this mixed distribution Mid-sampling yields the final control action This completes one online decision-making cycle.

[0143] Next, the system generates the final control action. At the same time, real-time risk signals and maximum similarity It performs continuous monitoring and multi-level threshold judgments, and triggers corresponding early warning and control actions accordingly. The early warning decision module will... and It compares the data with preset thresholds in real time and generates tiered warnings based on preset logic.

[0144] At the same time, the control execution module will output the final action of the strategy mixture. The commands are converted into specific, device-recognizable control commands (such as inverter power setpoint commands based on Modbus RTU or IEC 61850 protocols) and sent to the corresponding actuators (such as inverters and combiner boxes) of the photovoltaic power station through a secure communication link to complete early warning and control. The information is then reported to the staff through the operation and maintenance APP (application).

[0145] This application achieves proactive prevention. Through intrinsic algorithm modification, the intelligent agent has the ability to understand the early warning patterns of faults and can issue early warnings and take preventive control measures (such as proactive rate reduction) 30-60 seconds before the fault actually occurs. This fundamentally transforms the operation and maintenance mode from post-event remediation to proactive prevention, greatly reducing the risk of fault losses and equipment damage.

[0146] Meanwhile, it significantly improves the system's reliability and decision-making efficiency, eliminates complex external prediction models and removes prediction error propagation links, resulting in a simpler and more reliable system architecture. The parallelized decision-making architecture significantly reduces online inference latency (controllable within 200 milliseconds), meeting the real-time control requirements of photovoltaic power plants.

[0147] At the same time, enhance explainability and credibility, and improve the immediate risk factor. Similarity to history As a direct, quantifiable intermediate variable, it provides operations and maintenance personnel with clear criteria for risk assessment (how high the current risk is, and what types of failure precursors it resembles in the past). This interpretability greatly enhances operations and maintenance personnel's trust in the automated decision-making system.

[0148] It should be understood that the phrase "one embodiment" or "an 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, "in one embodiment" or "in an 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. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0149] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0150] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0151] The units described above as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected to achieve the purpose of the embodiments of this application according to actual needs. In addition, each functional unit in the embodiments of this application may be fully integrated into one processing unit, or each unit may be a separate unit, or two or more units may be integrated into one unit; the integrated unit may be implemented in hardware or in the form of hardware plus software functional units.

[0152] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause the device automatic test line to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0153] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined to obtain new method embodiments without conflict. The features disclosed in the several method or device embodiments provided in this application can be arbitrarily combined to obtain new method embodiments or device embodiments without conflict.

[0154] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A photovoltaic power plant operation and maintenance intelligent decision-making method based on multi-source data, applied to a photovoltaic power plant operation and maintenance intelligent decision-making system based on multi-source data, the system comprising a deep reinforcement learning agent, and a data acquisition module, a data preprocessing module, and an instruction execution module communicatively connected to the deep reinforcement learning agent, the deep reinforcement learning agent comprising a value network, a risk gating network, a precursor memory, a state enhancement module, a security network, and a policy generation module, the method comprising: Electrical and environmental data are acquired using the data acquisition module, and an original state vector is constructed based on the electrical and environmental data. The original state vector is standardized using the data preprocessing module to obtain a standardized state vector. The standardized state vector is used to extract features from the value network to obtain the current low-level detailed features and the current high-level semantic features. Using the aforementioned risk gating network, an instantaneous risk coefficient is determined based on the current high-level semantic features; Using the state enhancement module, the historical similarity between the current low-level detail feature and each historical low-level feature vector in the precursor memory is calculated. The maximum similarity in the historical similarity is concatenated with the standardized state vector to obtain the enhanced state vector. The historical low-level feature vector is the low-level feature vector of the time window of the precursor of the historical fault. Using the security network, a security feature vector is generated based on the enhanced state vector, the instantaneous risk coefficient, and the compressed low-level detail features obtained by projecting and compressing the current low-level detail features; based on the security feature vector, a security action distribution parameter containing the action mean and action variance is generated. Using the strategy generation module, control actions are generated based on the safety action distribution parameters; using the instruction execution module, a first control instruction is generated based on the control actions, and the first control instruction is sent to the execution mechanism for execution.

2. The method according to claim 1, characterized in that, The system also includes a basic network, and the method further includes: Using the aforementioned base network, basic action distribution parameters are generated based on the enhanced state vector; The step of generating control actions using the strategy generation module based on the security action distribution parameters includes: Using the strategy generation module, control actions are generated based on the basic action distribution parameters, the safety action distribution parameters, and the instantaneous risk coefficient.

3. The method according to claim 1, characterized in that, The system also includes an early warning decision module, and the method further includes: Using the aforementioned early warning decision module, when the immediate risk coefficient is greater than the first risk threshold and less than or equal to the second risk threshold, the early warning level is determined to be a state of concern, the current state is recorded, and a prompt message is output to the monitoring platform. Using the aforementioned early warning decision module, when the immediate risk coefficient is greater than the second risk threshold and less than or equal to the third risk threshold, or when the maximum similarity is greater than the first similarity threshold and less than or equal to the second similarity threshold, the early warning level is determined to be a high-risk state, a first early warning message containing risk information and suggested measures is generated, and the first early warning message is pushed to the operation and maintenance platform or mobile application. Using the aforementioned early warning decision module, when the immediate risk coefficient is greater than the third risk threshold and the maximum similarity is greater than the second similarity threshold, the early warning level is determined to be an emergency state, a second early warning message containing risk information and suggested measures is generated, a second control command to reduce output power is automatically executed according to the preset strategy configuration, and the second early warning message is pushed to the operation and maintenance platform or mobile application.

4. The method according to claim 1, characterized in that, The electrical data includes the total active power, total reactive power, and grid connection point voltage output from the AC side of the photovoltaic power station inverter. The environmental data includes ambient temperature and light intensity. Based on the electrical data and the environmental data, an original state vector is constructed, including: Based on the total active power, the total reactive power, the grid connection point voltage, the ambient temperature, and the light intensity, a five-dimensional original state vector is constructed; The step of standardizing the original state vector using the data preprocessing module to obtain a standardized state vector includes: Using the data preprocessing module, based on the first mean vector and the first standard deviation vector obtained from the statistical characteristics of the offline training dataset, linear transformations are performed on the data of each dimension of the original state vector to obtain a standardized state vector.

5. The method according to claim 1, characterized in that, The value network includes a shared feature layer and a high-level feature layer. The step of using the value network to extract features from the standardized state vector to obtain current low-level detailed features and current high-level semantic features includes: The normalized state vector is subjected to initial feature transformation and fusion using the fully connected layer of the shared feature layer, and the feature is activated using the ReLU activation function of the shared feature layer to obtain the current low-level detailed features; The current low-level detailed features are processed sequentially through the first fully connected layer, the ReLU activation function, the second fully connected layer, and the ReLU activation function of the high-level feature layer. Through layer-by-layer nonlinear transformation and feature activation, the current high-level semantic features are obtained.

6. The method according to claim 1, characterized in that, The method further includes: Acquire all confirmed actual fault events during the operation of the system; Mark the time point of each actual fault event; Based on the fault precursor time window, extract the historical low-level feature vector within the corresponding fault precursor time window before the fault occurrence time point. Each historical low-level feature vector is stored in the precursor memory in an ordered manner according to its timestamp.

7. The method according to claim 2, characterized in that, The step of generating control actions using the strategy generation module, based on the basic action distribution parameters, the safety action distribution parameters, and the immediate risk coefficient, includes: Based on the instantaneous risk coefficient, determine the first mixed weight of the basic action distribution parameter and the second mixed weight of the safe action distribution parameter; Based on the basic action distribution parameters, the first mixed weight, the safety action distribution parameters, and the second mixed weight, a mixed action distribution is generated; The control action is obtained by sampling from the mixed action distribution.

8. A photovoltaic power plant operation and maintenance intelligent decision-making system based on multi-source data, the system comprising a deep reinforcement learning agent, and a data acquisition module, a data preprocessing module, and an instruction execution module communicatively connected to the deep reinforcement learning agent, wherein the deep reinforcement learning agent comprises a value network, a risk gating network, a precursor memory, a state enhancement module, a security network, and a policy generation module, wherein: The data acquisition module is used to acquire electrical data and environmental data, and construct an original state vector based on the electrical data and environmental data; the data preprocessing module is used to standardize the original state vector to obtain a standardized state vector. The value network is used to extract features from the standardized state vector to obtain the current low-level detailed features and the current high-level semantic features. The risk gating network is used to determine the instantaneous risk coefficient based on the current high-level semantic features; The state enhancement module is used to calculate the historical similarity between the current low-level detail features and each historical low-level feature vector in the precursor memory, and to concatenate the maximum similarity in the historical similarity with the standardized state vector to obtain the enhanced state vector. The historical low-level feature vector is the low-level feature vector of the time window of the precursor of the historical fault. The security network is used to generate a security feature vector based on the enhanced state vector, the instantaneous risk coefficient, and the compressed low-level detail features obtained by projecting and compressing the current low-level detail features; and to generate a security action distribution parameter containing the action mean and action variance based on the security feature vector. The strategy generation module is used to generate control actions based on the security action distribution parameters; The instruction execution module is used to generate a first control instruction based on the control action and send the first control instruction to the execution mechanism for execution.

9. The system according to claim 8, characterized in that, The deep reinforcement learning agent also includes: A base network is used to generate basic action distribution parameters based on the enhanced state vector; The strategy generation module is used to generate control actions based on the basic action distribution parameters, the safety action distribution parameters, and the instantaneous risk coefficient.

10. The system according to claim 8, characterized in that, The system also includes: The early warning decision module is used to determine the early warning level as a state of concern when the instantaneous risk coefficient is greater than the first risk threshold and less than or equal to the second risk threshold, record the current status, and output prompt information on the monitoring platform. The early warning decision module is used to determine the early warning level as a high-risk state when the instantaneous risk coefficient is greater than the second risk threshold and less than or equal to the third risk threshold, or when the maximum similarity is greater than the first similarity threshold and less than or equal to the second similarity threshold, generate first early warning information containing risk information and suggested measures, and push the first early warning information to the operation and maintenance platform or mobile application. The early warning decision module is used to determine the early warning level as an emergency state when the immediate risk coefficient is greater than the third risk threshold and the maximum similarity is greater than the second similarity threshold, generate a second early warning message containing risk information and suggested measures, automatically execute a second control command to reduce output power according to a preset strategy configuration, and push the second early warning message to the operation and maintenance platform or mobile application.

Images