A photovoltaic power station intelligent supervision method based on digital twinning

Abstract

The application discloses a kind of photovoltaic field station wisdom supervision methods based on digital twinning, comprising the following steps: S1, obtains electric power time sequence signal and physical parameter sequence and pre-processes, forms training sample set;S2, constructs multiple expert DeepAR teacher model, generates future prediction path set;S3, construct light DeepAR student model and execute distillation training, obtain prediction model;S4, calculate prediction residual and combine defect confidence, sensor quality and digital twinning deviation form original evidence;S5, original evidence is mapped to fuzzy membership and input fuzzy Markov logic network, executes restricted causal reasoning;S6, output fault candidate node and causal path and by digital twinning simulation is verified;S7, parameter update is executed to prediction model and logic network, forms closed loop supervision mechanism.The application realizes the high-precision prediction of photovoltaic field station supervision, reliable fault inference and continuous adaptive optimization, and improves the operation safety and intelligent level.

Description

Technical Field

[0001] This invention relates to the field of digital operation and maintenance technology for new energy, and in particular to a smart monitoring method for photovoltaic power plants based on digital twins. Background Technology

[0002] With the rapid expansion of photovoltaic power generation, photovoltaic power plants face complex operating environments and diverse equipment types. Traditional operation and maintenance monitoring relies on experience-based judgment and static threshold detection, making it difficult to identify equipment faults and power anomalies in a timely and accurate manner. Existing prediction and diagnostic methods typically employ single time-series prediction models (such as ARIMA and LSTM) to model power generation. However, these methods cannot fully consider the impact of non-stationary factors such as weather conditions and irradiance variations, resulting in weak model generalization ability and insufficient prediction accuracy and stability.

[0003] Traditional models lack mechanisms for expressing and processing uncertain information when faced with abnormal sensor data or image detection noise, making them prone to misjudgment or omission. On the other hand, existing logical reasoning or causal analysis methods are mostly based on deterministic rules or Boolean logic, which cannot handle fuzzy evidence from multi-source heterogeneous data such as physical simulation, sensors, and image recognition. They also lack constrained reasoning mechanisms in complex causal chains, resulting in unreliable diagnostic results and low reasoning efficiency.

[0004] While digital twin technology has emerged for photovoltaic system modeling and simulation, existing digital twin systems often remain at the physical layer data mapping stage and have not yet achieved deep integration of data-driven prediction models and logical reasoning modules.

[0005] Therefore, how to provide a smart monitoring method for photovoltaic power plants based on digital twins is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a smart monitoring method for photovoltaic power plants based on digital twins. This invention combines multi-expert DeepAR prediction, knowledge distillation, fuzzy Markov logic networks, and digital twin simulation technology to achieve accurate prediction, reliable diagnosis, and closed-loop optimization of photovoltaic power anomalies.

[0007] A smart monitoring method for photovoltaic power plants based on digital twins according to an embodiment of the present invention includes the following steps:

[0008] S1. Obtain multi-channel power time-series signals and physical state parameters of photovoltaic power stations, perform wavelet denoising, normalization and sliding window segmentation processing to generate training sample sets and physical parameter sequences;

[0009] S2. Construct a multi-expert DeepAR teacher model, initialize K expert sub-models and gating network parameters, and perform supervised training based on the training sample set to obtain a trained DeepAR teacher model.

[0010] S3. Use the DeepAR teacher model to generate a set of future sample paths, construct the DeepAR student model and perform knowledge distillation training, and combine the physical consistency constraint loss function to obtain the distilled DeepAR student model.

[0011] S4. Deploy the DeepAR student model for online prediction, calculate the measurement power and prediction distribution residuals, and map the residuals, image defect confidence, sensor confidence and digital twin physical bias into a fuzzy evidence set.

[0012] S5. Construct a fuzzy Markov logic network, use fuzzy membership degree as predicate truth value, set rule weight, maximum causal chain depth and probability regularization parameters, and perform restricted inference based on fuzzy evidence set to output root cause candidates sorted by confidence.

[0013] S6. Perform simulation repair on the top N root cause candidates in the digital twin sandbox, calculate the power sample path and expected net benefit after repair, and select the optimal repair action.

[0014] S7. Based on the optimal repair action and root cause candidates, generate and issue work orders, and backfill the training sample set, DeepAR teacher model weights and fuzzy Markov logic network rule weights to complete closed-loop learning.

[0015] Optionally, S2 specifically includes:

[0016] S21. Based on the historical operation data of photovoltaic power stations, construct the expert classification criteria, and cluster the training sample set according to the clear sky irradiance mode, rapid cloud shadow change mode and weak irradiance mode to obtain the number of expert categories K and the sample index of each category.

[0017] S22. Based on the training sample set and physical parameter sequence, establish a corresponding DeepAR expert sub-model for each type of sample. The input format of the DeepAR expert sub-model is a unified input tensor formed by concatenating the power time series window sequence and the physical parameter sequence according to the timestamp. The output format is the conditional probability distribution parameters of the future prediction window.

[0018] S23. Specify the recurrent unit type as GRU for each DeepAR expert sub-model, specify the number of recurrent layers, the size of hidden units, and specify that the output layer contains only the distribution mean parameter and the distribution scale parameter.

[0019] S24. Construct a gating network. The input vector of the gating network is generated by concatenating the instantaneous values ​​of the irradiance, irradiance change rate and physical parameter sequence of the corresponding window in the training sample set. The output dimension of the gating network is K. Perform weighted synthesis on the output probability distribution of K expert sub-models.

[0020] S25. Initialize the parameters of K DeepAR expert sub-models and gated network parameters, load the sliding window samples of the training sample set in time order and perform forward propagation, and calculate the negative log-likelihood loss corresponding to the weighted synthesis output distribution.

[0021] S26. Based on the negative log-likelihood loss, perform gradient updates on the parameters of the K DeepAR expert sub-models and the gating network parameters, and repeat iterative training until the validation set loss satisfies the convergence condition.

[0022] S27. Save the DeepAR teacher model and the correspondence between expert categories and sample indices, and output the DeepAR teacher model.

[0023] Optionally, S3 specifically includes:

[0024] S31. Use the DeepAR teacher model to generate a sample path set for the future prediction window from the training sample set. The sample path set consists of multiple probability sampling sequences output by the DeepAR teacher model during the autoregressive prediction process.

[0025] S32. Construct a DeepAR student model. The DeepAR student model is a single-structure sequence prediction model. The input format of the DeepAR student model is generated by concatenating the power time series window sequence and the physical parameter sequence in the training sample set according to the timestamp to form a unified input tensor. The number of loop layers and the size of hidden units in the DeepAR student model are set to be smaller than the specifications of the number of loop layers and the size of hidden units in all DeepAR expert sub-models. The output format of the DeepAR student model is the conditional probability distribution parameters of the future prediction window.

[0026] S33. Initialize the recurrent unit type, number of recurrent layers, hidden unit size and output layer parameter dimension of the DeepAR student model, and load the sliding window samples corresponding to the sample path set and training sample set.

[0027] S34. Perform forward propagation on each batch of sliding window samples, calculate the conditional probability distribution parameters of the DeepAR student model output, and calculate the difference in probability distribution between the sample path set of the DeepAR teacher model and the student model sample path set.

[0028] S35. Construct the distillation loss, which consists of the distance term between the probability distribution of the sample path set output by the DeepAR teacher model and the probability distribution output by the DeepAR student model, as well as the physical consistency constraint term constructed from the sequence of physical parameters.

[0029] S36. Perform gradient updates on the DeepAR student model parameters based on the distillation loss, and repeat the forward propagation and gradient update until the distillation loss on the validation set meets the convergence criterion.

[0030] Optionally, S4 specifically includes:

[0031] S41. During the online inference process, obtain the latest timestamp data of the current measured power value and the physical parameter sequence, and concatenate the physical parameter sequence according to the timestamp into an online input tensor;

[0032] S42. Input the online input tensor into the DeepAR student model to generate the conditional probability distribution parameters of the future prediction window;

[0033] S43. Calculate the predicted residual sequence based on the conditional probability distribution parameters and the measured power value. The predicted residual sequence is obtained by subtracting the expected value of the student model output distribution from the measured power value by the timestamp.

[0034] S44. Collect the predicted residual sequence, the defect confidence level of the image detection output, the sensor quality assessment results, and the digital twin simulation bias, and arrange them by timestamp to form the original evidence sequence;

[0035] S45. Based on the predefined membership function, perform fuzzy mapping on the predicted residual sequence, defect confidence, sensor quality assessment result and digital twin simulation deviation in the original evidence sequence to generate residual fuzzy membership, defect fuzzy membership, quality fuzzy membership and physical consistency fuzzy membership.

[0036] S46. Combine the residual fuzzy membership degree, defect fuzzy membership degree, quality fuzzy membership degree and physical consistency fuzzy membership degree according to the timestamp to form a fuzzy evidence set.

[0037] Optionally, S45 specifically includes:

[0038] S451. Arrange the predicted residual sequence in the original evidence sequence into a one-dimensional real number sequence according to the timestamp, arrange the defect confidence of the image detection output into a one-dimensional normalized probability sequence according to the timestamp, arrange the sensor quality assessment results into a one-dimensional quality score sequence according to the timestamp, and arrange the digital twin simulation deviation into a one-dimensional physical deviation sequence according to the timestamp.

[0039] S452. Select the residual membership function from the predefined membership functions for the predicted residual sequence. The residual membership function is composed of a trapezoidal function. Input each timestamp value in the predicted residual sequence into the residual membership function to generate the corresponding residual fuzzy membership degree sequence.

[0040] S453. Select the defect membership function from the predefined membership functions for the defect confidence output of the image detection. The defect membership function is composed of monotonically increasing functions. Input each timestamp value in the defect confidence sequence into the defect membership function to generate the corresponding defect fuzzy membership sequence.

[0041] S454. Select the quality membership function from the predefined membership functions for the sensor quality assessment results. The quality membership function is composed of a monotonically decreasing function. Input each timestamp value in the sensor quality score sequence into the quality membership function to generate the corresponding quality fuzzy membership degree sequence.

[0042] S455. Select the physical consistency membership function from the predefined membership functions for the digital twin simulation deviation. The physical consistency membership function is composed of Gaussian functions. Input each timestamp value in the physical deviation sequence into the physical consistency membership function to generate the corresponding physical consistency fuzzy membership degree sequence.

[0043] S456. Save the residual fuzzy membership sequence, the defect fuzzy membership sequence, the quality fuzzy membership sequence, and the physical consistency fuzzy membership sequence as fuzzy membership sequences respectively.

[0044] Optionally, S5 specifically includes:

[0045] S51. Construct a fuzzy Markov logic network, and define a rule set, a predicate set, and a variable set. The rule set consists of weighted first-order logic clauses. The predicate set corresponds to the residual fuzzy membership degree, the defect fuzzy membership degree, the quality fuzzy membership degree, and the physical consistency fuzzy membership degree. The variable set corresponds to the timestamp and the device identifier.

[0046] S52. Initialize the rule weights in the fuzzy Markov logic network, set the maximum causal chain depth parameter and probability regularization coefficient, and limit the maximum length of the inference path.

[0047] S53. Input the fuzzy membership sequence into the fuzzy Markov logic network according to the timestamp as the fuzzy predicate truth value, calculate the weighted truth value of each clause according to the rule set, and calculate the rule confidence value by maximizing the posterior probability principle.

[0048] S54. Execute a restricted inference process, and use a heuristic search strategy to expand the inference chain downward from the starting node of the residual fuzzy membership degree in the rule set. Calculate the path confidence value and add a probability regularization term at each inference expansion. Terminate the expansion when the inference depth reaches the maximum length of the inference path or the path confidence value is lower than the preset threshold.

[0049] S55. Calculate the normalized confidence score for the terminal nodes of all inference paths and output the root cause candidates sorted by confidence score.

[0050] Optionally, S6 specifically includes:

[0051] S61. Input the root cause candidates into the digital twin simulation environment, and call the physical model of the corresponding device to generate simulation results;

[0052] S62. Compare the difference between the simulated output power and the measured power, and select the root cause candidate with the smallest error as the final root cause determination result based on the simulation error.

[0053] S63. Execute the corresponding maintenance actions based on the final root cause determination results, and synchronize the updated equipment parameters to the digital twin model.

[0054] Optionally, S7 specifically includes:

[0055] S71. After completing the operation and maintenance actions, obtain the updated equipment operation data and physical parameter sequence, compare the measured data with the simulation output of the digital twin model, and calculate the model error index.

[0056] S72. Perform joint updates on the parameters of the DeepAR student model and the fuzzy Markov logic network based on the model error index.

[0057] The beneficial effects of this invention are:

[0058] (1) By introducing a multi-expert DeepAR structure and a gating network mechanism, this invention can adaptively select the optimal expert model for different weather and irradiance patterns, thereby achieving high-precision prediction of photovoltaic power and effectively solving the problem of poor adaptability of traditional single models to non-stationary scenarios.

[0059] (2) This invention utilizes a knowledge distillation strategy combined with physical consistency constraints to transfer the prediction knowledge of the complex teacher model to the lightweight student model, which maintains the prediction accuracy and reduces the computational cost of the model, thus meeting the real-time requirements of online prediction for photovoltaic power plants.

[0060] (3) This invention integrates fuzzy Markov logic networks to achieve joint reasoning on multi-source fuzzy evidence such as residuals, defect confidence, sensor quality and physical deviations. It also improves reasoning efficiency and credibility through causal chain depth limitation and probability regularization, and finally realizes closed-loop optimization and intelligent decision-making in the digital twin environment. Attached Figure Description

[0061] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0062] Figure 1 This is a flowchart of a smart monitoring method for photovoltaic power plants based on digital twins proposed in this invention;

[0063] Figure 2 This is a structural diagram of a multi-expert DeepAR teacher model for a smart monitoring method of photovoltaic power plants based on digital twins proposed in this invention.

[0064] Figure 3 This is a diagram showing the fuzzy Markov logic network and digital twin closed-loop optimization structure of a smart monitoring method for photovoltaic power plants based on digital twins proposed in this invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0066] refer to Figure 1-3 A smart monitoring method for photovoltaic power plants based on digital twins includes the following steps:

[0067] S1. Obtain multi-channel power time-series signals and physical state parameters of photovoltaic power stations, perform wavelet denoising, normalization and sliding window segmentation processing to generate training sample sets and physical parameter sequences;

[0068] S2. Construct a multi-expert DeepAR teacher model, initialize K expert sub-models and gating network parameters, and perform supervised training based on the training sample set to obtain a trained DeepAR teacher model.

[0069] S3. Use the DeepAR teacher model to generate a set of future sample paths, construct the DeepAR student model and perform knowledge distillation training, and combine the physical consistency constraint loss function to obtain the distilled DeepAR student model.

[0070] S4. Deploy the DeepAR student model for online prediction, calculate the measurement power and prediction distribution residuals, and map the residuals, image defect confidence, sensor confidence and digital twin physical bias into a fuzzy evidence set.

[0071] S5. Construct a fuzzy Markov logic network, use fuzzy membership degree as predicate truth value, set rule weight, maximum causal chain depth and probability regularization parameters, and perform restricted inference based on fuzzy evidence set to output root cause candidates sorted by confidence.

[0072] S6. Perform simulation repair on the top N root cause candidates in the digital twin sandbox, calculate the power sample path and expected net benefit after repair, and select the optimal repair action.

[0073] S7. Based on the optimal repair action and root cause candidates, generate and issue work orders, and backfill the training sample set, DeepAR teacher model weights and fuzzy Markov logic network rule weights to complete closed-loop learning.

[0074] In this embodiment, S1 specifically includes:

[0075] S11. The photovoltaic power station deploys various types of sensor devices, including string current and voltage acquisition equipment, inverter operation information acquisition equipment, and irradiance and environmental parameter monitoring instruments. These sensor devices periodically output multi-channel time-series data reflecting the current operating status. In this implementation, these data are first aligned with a unified timestamp, and all acquisition channels are arranged according to the same timeline.

[0076] S12. After time alignment is completed, noise suppression processing is performed on the original time-series signal of the power acquisition channel. Noise suppression adopts wavelet denoising, and by performing multi-scale decomposition, thresholding, and reconstruction operations on the signal, a power time-series signal with significantly reduced noise is obtained; the denoised signal is normalized according to a preset dimension to ensure that data of different dimensions are consistent in numerical scale.

[0077] S13. When constructing time-series structured samples for training, the continuous time series is segmented using a sliding window method. The sliding window length and step size are set according to the typical power change cycle in the photovoltaic scenario. In this implementation, each window segment is regarded as an independent sample, and fields containing power time series, irradiance, component temperature and wind speed are extracted from it to form a training sample set.

[0078] S14. Synchronously generate physical parameter sequences; the physical parameter sequences correspond one-to-one with the power time series data according to the timestamp.

[0079] In this embodiment, S2 specifically includes:

[0080] S21. Based on the historical operation data of photovoltaic power stations, construct the expert classification criteria, and cluster the training sample set according to the clear sky irradiance mode, rapid cloud shadow change mode and weak irradiance mode to obtain the number of expert categories K and the sample index of each category.

[0081] S22. Based on the training sample set and physical parameter sequence, establish a corresponding DeepAR expert sub-model for each type of sample. The input format of the DeepAR expert sub-model is a unified input tensor formed by concatenating the power time series window sequence and the physical parameter sequence according to the timestamp. The output format is the conditional probability distribution parameters of the future prediction window.

[0082] S23. Specify the recurrent unit type as GRU for each DeepAR expert sub-model, specify the number of recurrent layers, the size of hidden units, and specify that the output layer contains only the distribution mean parameter and the distribution scale parameter.

[0083] S24. Construct a gating network. The input vector of the gating network is generated by concatenating the instantaneous values ​​of the irradiance, irradiance change rate and physical parameter sequence of the corresponding window in the training sample set. The output dimension of the gating network is K. Perform weighted synthesis on the output probability distribution of K expert sub-models.

[0084] S25. Initialize the parameters of K DeepAR expert sub-models and gated network parameters, load the sliding window samples of the training sample set in time order and perform forward propagation, and calculate the negative log-likelihood loss corresponding to the weighted synthesis output distribution.

[0085] S26. Based on the negative log-likelihood loss, perform gradient updates on the parameters of the K DeepAR expert sub-models and the gated network parameters, and repeat iterative training until the validation set loss meets the convergence condition. The convergence condition is that when the change in the negative log-likelihood loss of the validation set is less than 0.0001 for three consecutive iterations, it is considered that the loss has entered the stable interval, and training stops. The maximum number of rounds is set to 200.

[0086] S27. Save the DeepAR teacher model and the correspondence between expert categories and sample indices, and output the DeepAR teacher model.

[0087] In this embodiment, S21 specifically includes:

[0088] S211. Extract key variables from historical data to characterize the irradiance state, including instantaneous irradiance, rate of change of irradiance per unit time, and amplitude of irradiance fluctuation within a continuous time window; instantaneous irradiance reflects the intensity of light, the rate of change reflects the rapid changes caused by cloud shadows, and the amplitude of fluctuation reflects the irradiance disturbance under unstable weather conditions; in this implementation, the extracted irradiance variables are organized according to a unified timestamp to form a sequence of irradiance feature vectors for cluster analysis.

[0089] S212. An unsupervised clustering method is used to classify the irradiance feature vectors. In this implementation, a distance-based clustering algorithm is selected to divide the irradiance feature vectors of all time windows into three categories: clear sky irradiance mode, rapid cloud shadow change mode, and weak irradiance mode. The clear sky mode corresponds to the situation where the irradiance is stable and changes slowly. The rapid cloud shadow change mode corresponds to the situation where the irradiance rises or falls rapidly in a short period of time, which is determined by setting a threshold. The weak irradiance mode corresponds to the low irradiance level on cloudy days or in winter. After the clustering algorithm is executed, the category label of each time window is obtained.

[0090] S213. After clustering is completed, this embodiment maps each category label to an expert category index and forms three sample sets based on the time window index; each expert sub-model corresponds to an irradiation pattern, and each sample set serves as the training data source for the corresponding expert sub-model; this embodiment ultimately determines the number of expert categories K to be 3 through this step, and records the sample time window index corresponding to each category.

[0091] In this embodiment, the specification of the loop layer number and hidden unit size in S23 specifically includes:

[0092] S231. The number of loop layers is limited to between 1 and 3. In the clear-sky irradiance mode, the absolute value of the irradiance change rate under clear-sky conditions is mostly less than 20 W / m²·min. This implementation considers this mode as a low-change mode. Verification of the concentration error comparison shows that setting the number of loop layers to 1 is sufficient to fit the temporal characteristics of this mode. In the rapid cloud shadow change mode, the absolute value of the irradiance change rate exceeds 60 W / m²·min, and the irradiance fluctuation amplitude within five consecutive minutes exceeds 150 W / m². This mode exhibits the most significant temporal fluctuations; therefore, the number of loop layers is set to 3. In the weak irradiance mode, the irradiance is often below 200 W / m², and the absolute value of the change rate is between 0 and 40 W / m²·min. To balance temporal stability and model expressiveness, this implementation sets the number of loop layers to 2.

[0093] S232. The hidden unit size is selected between 32 and 128 units, based on the complexity of the irradiance characteristics and the sample size range. In the clear-sky irradiance mode, the sample fluctuation range for most days is less than 5%, where the sample fluctuation range refers to the ratio of the maximum power value to the standard deviation of the time window. Verification shows that setting the hidden unit size to 64 units in this implementation achieves the optimal balance between error stability and training time when the sample size is sufficient for more than 30 days. In the rapid cloud shadow change mode, the power change amplitude within the time window exceeds 15%, which is a high-dynamic mode; in this implementation, the hidden unit size is set to 128 units. In the weak irradiance mode, the irradiance is consistently within the range of 0–200 W / m², and the power change amplitude typically does not exceed 3%; in this implementation, the hidden unit size is set to 32 units.

[0094] In this embodiment, the gating network in S24 specifically includes:

[0095] S241. The first hidden layer uses a fully connected structure with 64 hidden units and ReLU activation function.

[0096] S242. The second hidden layer is a fully connected structure with 32 hidden units and ReLU activation function.

[0097] S243. The output layer is a fully connected layer with an output dimension of K, which is the same as the number of expert sub-models. The output layer activation function is Softmax.

[0098] S244. The weight vector output by the gated network is weighted and synthesized with the conditional probability distribution parameters of each expert sub-model.

[0099] In this embodiment, S3 specifically includes:

[0100] S31. Use the DeepAR teacher model to generate a sample path set for the future prediction window from the training sample set. The sample path set consists of multiple probability sampling sequences output by the DeepAR teacher model during the autoregressive prediction process.

[0101] S32. Construct a DeepAR student model. The DeepAR student model is a single-structure sequence prediction model. The input format of the DeepAR student model is generated by concatenating the power time series window sequence and the physical parameter sequence in the training sample set according to the timestamp to form a unified input tensor. The number of loop layers and the size of hidden units in the DeepAR student model are set to be smaller than the specifications of the number of loop layers and the size of hidden units in all DeepAR expert sub-models. The output format of the DeepAR student model is the conditional probability distribution parameters of the future prediction window.

[0102] S33. Initialize the recurrent unit type, number of recurrent layers, hidden unit size and output layer parameter dimension of the DeepAR student model, and load the sliding window samples corresponding to the sample path set and training sample set.

[0103] S34. Perform forward propagation on each batch of sliding window samples, calculate the conditional probability distribution parameters of the DeepAR student model output, and calculate the difference in probability distribution between the sample path set of the DeepAR teacher model and the student model sample path set.

[0104] S35. Construct the distillation loss, which consists of the distance term between the probability distribution of the sample path set output by the DeepAR teacher model and the probability distribution output by the DeepAR student model, as well as the physical consistency constraint term constructed from the sequence of physical parameters.

[0105] S36. Perform gradient updates on the DeepAR student model parameters based on the distillation loss, and repeat the forward propagation and gradient update until the validation set distillation loss meets the convergence criterion. The convergence criterion is to stop training when the absolute value of the difference between the validation set distillation loss and the two hundredth training rounds is less than 0.001.

[0106] In this embodiment, S31 specifically includes:

[0107] S311. After receiving input, the DeepAR teacher model completes a forward calculation through an internal loop structure to obtain the conditional probability distribution of the first future time point. This implementation method performs a sampling operation based on the probability distribution to obtain a specific predicted value.

[0108] S312. The predicted value obtained by sampling at the previous time point is directly used as the power input at the current time point. The model loads the physical parameter data corresponding to this time point and recombines it into the input vector for the next step according to the input format of the DeepAR teacher model.

[0109] S313. Following S311 and S312, the rule of "using the predicted value of the previous time point as the input of the next time point" is continuously advanced, generating predicted values ​​sequentially at future time points. Once the entire future window has been predicted, a complete prediction path is formed. Since each prediction step obtains values ​​through probability sampling, the model can repeatedly execute the autoregressive process under the same input conditions, forming multiple prediction paths. All paths together constitute a sample path set used for distillation training.

[0110] In this embodiment, S35 specifically includes:

[0111] S351. Compare the teacher-side reference distribution with the probability distribution output by the student model at each time point, calculate the distribution difference at each time point, and accumulate this difference over the entire future prediction window to form a loss term that characterizes the "student prediction distribution approximates the teacher prediction distribution".

[0112] S352. The second part of the distillation loss is used to measure the physical consistency deviation between the student model output and the physical parameters. In the specific calculation process, for each sample at each time point within the future prediction window, the predicted power output by the student model at that time point is first taken out, and then the sequence of physical parameters generated by the digital twin model at the same time point is taken out. According to the physical laws of photovoltaic power generation, the target relationship between the predicted power value and the physical parameters is constructed. At each time point, the deviation between the student prediction result and these target physical relationships is calculated respectively, and the deviation is accumulated within the entire prediction window to form a physical consistency constraint term.

[0113] S353. Add the distribution difference term and the physical consistency constraint term according to the preset weights to form the complete distillation loss.

[0114] In this embodiment, S4 specifically includes:

[0115] S41. During the online inference process, obtain the latest timestamp data of the current measured power value and the physical parameter sequence, and concatenate the physical parameter sequence according to the timestamp into an online input tensor;

[0116] S42. Input the online input tensor into the DeepAR student model to generate the conditional probability distribution parameters of the future prediction window;

[0117] S43. Calculate the predicted residual sequence based on the conditional probability distribution parameters and the measured power value. The predicted residual sequence is obtained by subtracting the expected value of the student model output distribution from the measured power value by the timestamp.

[0118] S44. Collect the predicted residual sequence, the defect confidence level of the image detection output, the sensor quality assessment results, and the digital twin simulation bias, and arrange them by timestamp to form the original evidence sequence;

[0119] S45. Based on the predefined membership function, perform fuzzy mapping on the predicted residual sequence, defect confidence, sensor quality assessment result and digital twin simulation deviation in the original evidence sequence to generate residual fuzzy membership, defect fuzzy membership, quality fuzzy membership and physical consistency fuzzy membership.

[0120] S46. Combine the residual fuzzy membership degree, defect fuzzy membership degree, quality fuzzy membership degree and physical consistency fuzzy membership degree according to the timestamp to form a fuzzy evidence set.

[0121] In this embodiment, S44 specifically includes:

[0122] The defect confidence score output by the image detection is based on existing image recognition technology and is generated by existing intelligent detection algorithms for photovoltaic modules. Fixed camera equipment is deployed on-site to collect visible light and infrared images of the photovoltaic module surface, and these images are input into existing target detection networks or surface defect recognition networks. These networks output the defect probability corresponding to each image region based on hot spots, cracks, occlusions, dust accumulation, or other abnormal image features on the module surface, and provide the defect confidence score in the form of a probability value. In this implementation, the confidence score output by the detection algorithm is directly used as the image detection evidence for that timestamp.

[0123] The sensor quality assessment results are based on existing data integrity and stability detection technologies and are generated by existing sensor health monitoring methods. These methods perform stability and integrity checks on the real-time outputs of current sensors, voltage sensors, and irradiance sensors in photovoltaic power plants, including output continuity, numerical fluctuation range, noise amplitude, and signal drift. If a sensor experiences packet loss, abrupt changes, drift, or abnormal fluctuations, existing sensor quality assessment algorithms calculate a corresponding quality score based on the degree of deviation; this implementation directly uses this score as the sensor quality assessment result.

[0124] The digital twin simulation deviation is based on existing digital twin simulation technology. The simulation system uses physical parameters, component models, inverter models and control strategies consistent with the real equipment to simulate the theoretical power generation of a photovoltaic power station under the current environmental conditions. By comparing the simulated power at the same time stamp with the actual measured power, a deviation value used to characterize the simulation error can be obtained. This deviation is further input into the evidence sequence of this embodiment as the digital twin simulation deviation. This embodiment does not modify the simulation algorithm itself, but directly uses the output of the existing digital twin system as input.

[0125] In this embodiment, S45 specifically includes:

[0126] S451. Arrange the predicted residual sequence in the original evidence sequence into a one-dimensional real number sequence according to the timestamp, arrange the defect confidence of the image detection output into a one-dimensional normalized probability sequence according to the timestamp, arrange the sensor quality assessment results into a one-dimensional quality score sequence according to the timestamp, and arrange the digital twin simulation deviation into a one-dimensional physical deviation sequence according to the timestamp.

[0127] S452. Select the residual membership function from the predefined membership functions for the predicted residual sequence. The residual membership function is composed of a trapezoidal function. Input each timestamp value in the predicted residual sequence into the residual membership function to generate the corresponding residual fuzzy membership degree sequence.

[0128] S453. Select the defect membership function from the predefined membership functions for the defect confidence output of the image detection. The defect membership function is composed of monotonically increasing S-type membership functions. Input each timestamp value in the defect confidence sequence into the defect membership function to generate the corresponding defect fuzzy membership sequence.

[0129] S454. Select the quality membership function from the predefined membership functions for the sensor quality assessment results. The quality membership function is composed of monotonically decreasing S-shaped membership functions. Input each timestamp value in the sensor quality score sequence into the quality membership function to generate the corresponding quality fuzzy membership degree sequence.

[0130] S455. Select the physical consistency membership function from the predefined membership functions for the digital twin simulation deviation. The physical consistency membership function is composed of Gaussian functions. Input each timestamp value in the physical deviation sequence into the physical consistency membership function to generate the corresponding physical consistency fuzzy membership degree sequence.

[0131] S456. Save the residual fuzzy membership sequence, the defect fuzzy membership sequence, the quality fuzzy membership sequence, and the physical consistency fuzzy membership sequence as fuzzy membership sequences respectively.

[0132] In this embodiment, S5 specifically includes:

[0133] S51. Construct a fuzzy Markov logic network, and define a rule set, a predicate set, and a variable set. The rule set consists of weighted first-order logic clauses. The predicate set corresponds to the residual fuzzy membership degree, the defect fuzzy membership degree, the quality fuzzy membership degree, and the physical consistency fuzzy membership degree. The variable set corresponds to the timestamp and the device identifier.

[0134] S52. Initialize the rule weights in the fuzzy Markov logic network, set the maximum causal chain depth parameter and probability regularization coefficient, and limit the maximum length of the inference path.

[0135] S53. Input the fuzzy membership sequence into the fuzzy Markov logic network according to the timestamp as the fuzzy predicate truth value, calculate the weighted truth value of each clause according to the rule set, and calculate the rule confidence value by maximizing the posterior probability principle.

[0136] S54. Execute a restricted inference process, and use a heuristic search strategy to expand the inference chain downward from the starting node of the residual fuzzy membership degree in the rule set. Calculate the path confidence value and add a probability regularization term at each inference expansion. Terminate the expansion when the inference depth reaches the maximum length of the inference path or the path confidence value is lower than the preset threshold.

[0137] S55. Calculate the normalized confidence score for the terminal nodes of all inference paths and output the root cause candidates sorted by confidence score.

[0138] In this embodiment, S52 sets a maximum length for the inference path, based on the following:

[0139] The causal propagation hierarchy between photovoltaic power station equipment typically does not exceed five levels: modules, strings, inverters, combiner zones, and grid-connected nodes. Paths exceeding this range have no physical meaning. Statistical analysis of historical fault events shows that the causal chain length of typical faults is concentrated at levels two to three, while the longest actual propagation chain does not exceed level five. When performing parameter perturbation simulations in a digital twin environment, power changes can be fully reflected within the first three to five causal nodes. To ensure the executability of Markov logic network reasoning in online scenarios, the reasoning depth needs to be controlled. Based on the comprehensive considerations of the above physical, statistical, simulation, and engineering efficiency, this implementation method sets the maximum length of the reasoning path to no more than five levels.

[0140] In this embodiment, S53 specifically includes:

[0141] Regarding the fuzzy operation rules, this implementation adopts the t-norm operation in existing fuzzy logic as the conjunction calculation method inside the clause, specifically using the Łukasiewicz t-norm.

[0142] In this embodiment, the S54 regularization term specifically includes:

[0143] By statistically analyzing historical causal relationship samples from photovoltaic power plants, the distribution of rule weight values ​​in different operating scenarios can be obtained, which can be used to construct the regularization term of the rule weight.

[0144] By analyzing the causal chain length distribution and node transition probability distribution generated by digital twin simulation, we can obtain the probability reference range of the inference path at different depths, which can be used to construct path depth-related regularization terms.

[0145] By statistically analyzing the variance of the fuzzy membership sequence across different evidence types, a reference interval for membership variation can be obtained, which can be used to construct a regularization term for membership smoothness.

[0146] In this embodiment, S6 specifically includes:

[0147] S61. Input the root cause candidates into the digital twin simulation environment, and call the physical model of the corresponding device to generate simulation results;

[0148] S62. Compare the difference between the simulated output power and the measured power, and select the root cause candidate with the smallest error as the final root cause determination result based on the simulation error.

[0149] S63. Execute the corresponding maintenance actions based on the final root cause determination results, and synchronize the updated equipment parameters to the digital twin model.

[0150] In this embodiment, S7 specifically includes:

[0151] S71. After completing the operation and maintenance actions, obtain the updated equipment operation data and physical parameter sequence, compare the measured data with the simulation output of the digital twin model, and calculate the model error index; the error index consists of the difference between the measured data and the digital twin simulation output.

[0152] S72. Perform joint updates on the parameters of the DeepAR student model and the fuzzy Markov logic network based on the model error index; obtain the latest measured data after the operation and maintenance actions are performed, and calculate the error index between this batch of data and the digital twin simulation results; input the error index into the DeepAR student model and the fuzzy Markov logic network respectively, and perform gradient updates on the recurrent layer parameters and output layer parameters of the DeepAR student model based on the error index, and perform numerical updates on the rule weights, fuzzy membership parameters and path probabilities of the fuzzy Markov logic network based on the error index.

[0153] Example 1:

[0154] To verify the feasibility of this invention in practice, it was applied to a photovoltaic power generation scenario with a typical string-inverter-grid multi-stage structure. This scenario includes a large number of photovoltaic modules, strings, and multiple inverters. During operation, the power output exhibits significant nonlinearity and uncertainty due to factors such as changes in irradiance, local shading, temperature fluctuations, and equipment aging. Traditional monitoring methods often struggle to predict power changes in a timely manner or accurately locate potential fault sources when dealing with rapid changes in illumination or minor equipment degradation, leading to a decline in operation and maintenance quality. Therefore, a smart monitoring method is needed that can integrate multi-source data, possess causal inference capabilities, and enable verification and closed-loop updates through digital twins.

[0155] During implementation, the historical operating data of the photovoltaic power station was first organized, and parameters such as module current, voltage, inverter efficiency, temperature, irradiance, and wind speed were aligned by timestamps to construct a multi-channel time-series signal set. This data was then denoised, normalized, and sliced ​​into fixed windows to generate a data sequence for training the DeepAR model. Simultaneously, physical parameters from the same batch were input into the digital twin model for use as simulation references in subsequent processes.

[0156] In the prediction model construction phase, this embodiment divides all training samples into three categories based on the differences in irradiance patterns: clear sky irradiance, rapid cloud shadow change, and weak irradiance. A corresponding DeepAR expert sub-model is trained for each category, enabling the temporal patterns corresponding to different patterns to be expressed separately. After completing the teacher model training, a large number of probability sampling paths are generated from it to train a lightweight student model, allowing the student model to achieve prediction capabilities close to those of the teacher model at a lower computational cost.

[0157] In the process of intelligent supervision, real-time measured power data and physical parameter data are accessed and input into a student model to obtain predictions of future power distribution. The expected value of the predicted distribution is compared with the measured power to obtain a point-by-point prediction residual sequence. Simultaneously, image detection generates defect confidence scores for component surface conditions, sensor quality scores are obtained through sensor diagnostic algorithms, and the digital twin model provides simulation bias information. This information is organized into a sequence of original evidence by timestamps.

[0158] Subsequently, using preset membership functions, the prediction residuals, defect confidence scores, sensor quality scores, and digital twin biases are converted into fuzzy membership degrees. Trapezoidal membership functions are used for residual mapping, sigmoid membership functions for defect confidence and quality scores, and Gaussian functions for digital twin bias, forming a fuzzy evidence set. This set is input into a fuzzy MLN, and by limiting the maximum inference path length to no more than 5 and employing a regularization term based on data statistics and simulation results, causal inference is completed, and possible fault nodes and corresponding causal chains are output.

[0159] After identifying candidate faults, the digital twin model performs simulation verification, comparing the simulation output with actual data to determine if the causal chain matches the actual situation. When inconsistencies arise, the weight parameters of the DeepAR student model and the rule weights of the fuzzy MLN are jointly updated using error metrics, thus completing the closed loop of intelligent supervision.

[0160] To verify the effectiveness of this method, this embodiment compares the performance of traditional methods and the method of this invention on several key indicators. Experimental results show that in power prediction over the next 15 minutes, the average prediction error of the traditional method is 6.8%, while that of this method is 3.1%; the peak error of the traditional method is 12.4%, while that of this method is 5.6%. Regarding fault location deviation, the traditional method has an average deviation of 2.3 nodes, while this method has a deviation of 0.8 nodes. In terms of consistency rate in causal chain inference, this method achieves 89.5%, while the traditional method achieves 71.2%. These data demonstrate that the method of this invention significantly outperforms traditional methods in terms of prediction accuracy, location accuracy, and causal inference capability.

[0161] Table 1: Performance Comparison Data of Photovoltaic Smart Monitoring Methods

[0162] index Traditional methods Method of the present invention Increase Average power prediction error (%) 6.8 3.1 3.7 Peak error (%) 12.4 5.6 6.8 Average deviation of fault location (number of nodes) 2.3 0.8 1.5 Consistency rate of causal chain inference (%) 71.2 89.5 18.3

[0163] The data shown in Table 1 demonstrates that the method of this invention significantly outperforms traditional methods in several key indicators. Regarding the average power prediction error, the method's 3.1% is more than 50% lower than the traditional method's 6.8%, indicating that the multi-expert DeepAR and distillation strategies employed can more accurately characterize the dynamic characteristics of photovoltaic output. As for the peak error, the method's 5.6% is lower than the traditional value of 12.4%, demonstrating that the invention's prediction capability is more stable under drastically changing conditions.

[0164] In fault location tasks, the average deviation of this method is 0.8 nodes, while that of the traditional method is 2.3 nodes, a difference of more than 1.5 nodes, enabling maintenance personnel to pinpoint the source of the problem more quickly. Regarding the consistency rate of causal chain inference, this method achieves 89.5%, an improvement of 18.3% compared to the traditional method, indicating that fuzzy MLN combined with digital twin verification can more accurately reconstruct the true causal structure and obtain more reliable inference results.

[0165] In summary, the method of this invention can improve real-time prediction, fault identification, causal inference and digital twin verification, and has the value of being promoted and applied in photovoltaic smart supervision scenarios.

[0166] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart monitoring method for photovoltaic power plants based on digital twins, characterized in that, Includes the following steps: S1. Obtain multi-channel power time-series signals and physical state parameters of photovoltaic power stations, perform wavelet denoising, normalization and sliding window segmentation processing to generate training sample sets and physical parameter sequences; S2. Construct a multi-expert DeepAR teacher model. Based on the historical operation data of photovoltaic power stations, construct the expert division criteria. Cluster the training sample set according to the clear sky irradiance mode, rapid cloud shadow change mode and weak irradiance mode to obtain the number of expert categories K. Initialize K expert sub-models and gating network parameters, and perform supervised training based on the training sample set to obtain the trained DeepAR teacher model. S3. Use the DeepAR teacher model to generate a set of future sample paths, construct the DeepAR student model and perform knowledge distillation training, and combine the physical consistency constraint loss function to obtain the distilled DeepAR student model. S4. Deploy the DeepAR student model for online prediction, calculate the measurement power and prediction distribution residuals, and map the residuals, image defect confidence, sensor confidence and digital twin physical bias into a fuzzy evidence set. S5. Construct a fuzzy Markov logic network, use fuzzy membership degree as predicate truth value, set rule weight, maximum causal chain depth and probability regularization parameters, and perform restricted inference based on fuzzy evidence set to output root cause candidates sorted by confidence. S6. Perform simulation repair on the top N root cause candidates in the digital twin sandbox, calculate the power sample path and expected net benefit after repair, and select the optimal repair action. S7. Based on the optimal repair action and root cause candidates, generate and issue work orders, and backfill the training sample set, DeepAR teacher model weights and fuzzy Markov logic network rule weights to complete closed-loop learning.

2. The intelligent monitoring method for photovoltaic power plants based on digital twins according to claim 1, characterized in that, S2 specifically includes: S21. Based on the historical operation data of photovoltaic power stations, construct the expert classification criteria, and cluster the training sample set according to the clear sky irradiance mode, rapid cloud shadow change mode and weak irradiance mode to obtain the number of expert categories K and the sample index of each category. S22. Based on the training sample set and physical parameter sequence, establish a corresponding DeepAR expert sub-model for each type of sample. The input format of the DeepAR expert sub-model is a unified input tensor formed by concatenating the power time series window sequence and the physical parameter sequence according to the timestamp. The output format is the conditional probability distribution parameters of the future prediction window. S23. Specify the recurrent unit type as GRU for each DeepAR expert sub-model, specify the number of recurrent layers, the size of hidden units, and specify that the output layer contains only the distribution mean parameter and the distribution scale parameter. S24. Construct a gating network. The input vector of the gating network is generated by concatenating the instantaneous values ​​of the irradiance, irradiance change rate and physical parameter sequence of the corresponding window in the training sample set. The output dimension of the gating network is K. Perform weighted synthesis on the output probability distribution of K expert sub-models. S25. Initialize the parameters of K DeepAR expert sub-models and gated network parameters, load the sliding window samples of the training sample set in time order and perform forward propagation, and calculate the negative log-likelihood loss corresponding to the weighted synthesis output distribution. S26. Based on the negative log-likelihood loss, perform gradient updates on the parameters of the K DeepAR expert sub-models and the gating network parameters, and repeat iterative training until the validation set loss satisfies the convergence condition. S27. Save the DeepAR teacher model and the correspondence between expert categories and sample indices, and output the DeepAR teacher model.

3. The intelligent monitoring method for photovoltaic power plants based on digital twins according to claim 2, characterized in that, S3 specifically includes: S31. Use the DeepAR teacher model to generate a sample path set for the future prediction window from the training sample set. The sample path set consists of multiple probability sampling sequences output by the DeepAR teacher model during the autoregressive prediction process. S32. Construct a DeepAR student model. The DeepAR student model is a single-structure sequence prediction model. The input format of the DeepAR student model is generated by concatenating the power time series window sequence and the physical parameter sequence in the training sample set according to the timestamp to form a unified input tensor. The number of loop layers and the size of hidden units in the DeepAR student model are set to be smaller than the specifications of the number of loop layers and the size of hidden units in all DeepAR expert sub-models. The output format of the DeepAR student model is the conditional probability distribution parameters of the future prediction window. S33. Initialize the recurrent unit type, number of recurrent layers, hidden unit size and output layer parameter dimension of the DeepAR student model, and load the sliding window samples corresponding to the sample path set and training sample set. S34. Perform forward propagation on each batch of sliding window samples, calculate the conditional probability distribution parameters of the DeepAR student model output, and calculate the difference in probability distribution between the sample path set of the DeepAR teacher model and the student model sample path set. S35. Construct the distillation loss, which consists of the distance term between the probability distribution of the sample path set output by the DeepAR teacher model and the probability distribution output by the DeepAR student model, as well as the physical consistency constraint term constructed from the sequence of physical parameters. S36. Perform gradient updates on the DeepAR student model parameters based on the distillation loss, and repeat the forward propagation and gradient update until the distillation loss on the validation set meets the convergence criterion.

4. The intelligent monitoring method for photovoltaic power plants based on digital twins according to claim 3, characterized in that, S4 specifically includes: S41. During the online inference process, obtain the latest timestamp data of the current measured power value and the physical parameter sequence, and concatenate the physical parameter sequence according to the timestamp into an online input tensor; S42. Input the online input tensor into the DeepAR student model to generate the conditional probability distribution parameters of the future prediction window; S43. Calculate the predicted residual sequence based on the conditional probability distribution parameters and the measured power value. The predicted residual sequence is obtained by subtracting the expected value of the student model output distribution from the measured power value by the timestamp. S44. Collect the predicted residual sequence, the defect confidence level of the image detection output, the sensor quality assessment results, and the digital twin simulation bias, and arrange them by timestamp to form the original evidence sequence; S45. Based on the predefined membership function, perform fuzzy mapping on the predicted residual sequence, defect confidence, sensor quality assessment result and digital twin simulation deviation in the original evidence sequence to generate residual fuzzy membership, defect fuzzy membership, quality fuzzy membership and physical consistency fuzzy membership. S46. Combine the residual fuzzy membership degree, defect fuzzy membership degree, quality fuzzy membership degree and physical consistency fuzzy membership degree according to the timestamp to form a fuzzy evidence set.

5. The intelligent monitoring method for photovoltaic power plants based on digital twins according to claim 4, characterized in that, Specifically, S45 includes: S451. Arrange the predicted residual sequence in the original evidence sequence into a one-dimensional real number sequence according to the timestamp, arrange the defect confidence of the image detection output into a one-dimensional normalized probability sequence according to the timestamp, arrange the sensor quality assessment results into a one-dimensional quality score sequence according to the timestamp, and arrange the digital twin simulation deviation into a one-dimensional physical deviation sequence according to the timestamp. S452. Select the residual membership function from the predefined membership functions for the predicted residual sequence. The residual membership function is composed of a trapezoidal function. Input each timestamp value in the predicted residual sequence into the residual membership function to generate the corresponding residual fuzzy membership degree sequence. S453. Select the defect membership function from the predefined membership functions for the defect confidence output of the image detection. The defect membership function is composed of monotonically increasing functions. Input each timestamp value in the defect confidence sequence into the defect membership function to generate the corresponding defect fuzzy membership sequence. S454. Select the quality membership function from the predefined membership functions for the sensor quality assessment results. The quality membership function is composed of a monotonically decreasing function. Input each timestamp value in the sensor quality score sequence into the quality membership function to generate the corresponding quality fuzzy membership degree sequence. S455. Select the physical consistency membership function from the predefined membership functions for the digital twin simulation deviation. The physical consistency membership function is composed of Gaussian functions. Input each timestamp value in the physical deviation sequence into the physical consistency membership function to generate the corresponding physical consistency fuzzy membership degree sequence. S456. Save the residual fuzzy membership sequence, the defect fuzzy membership sequence, the quality fuzzy membership sequence, and the physical consistency fuzzy membership sequence as fuzzy membership sequences respectively.

6. The intelligent monitoring method for photovoltaic power plants based on digital twins according to claim 4, characterized in that, S5 specifically includes: S51. Construct a fuzzy Markov logic network, and define a rule set, a predicate set, and a variable set. The rule set consists of weighted first-order logic clauses. The predicate set corresponds to the residual fuzzy membership degree, the defect fuzzy membership degree, the quality fuzzy membership degree, and the physical consistency fuzzy membership degree. The variable set corresponds to the timestamp and the device identifier. S52. Initialize the rule weights in the fuzzy Markov logic network, set the maximum causal chain depth parameter and probability regularization coefficient, and limit the maximum length of the inference path. S53. Input the fuzzy membership sequence into the fuzzy Markov logic network according to the timestamp as the fuzzy predicate truth value, calculate the weighted truth value of each clause according to the rule set, and calculate the rule confidence value by maximizing the posterior probability principle. S54. Execute a restricted inference process, and use a heuristic search strategy to expand the inference chain downward from the starting node of the residual fuzzy membership degree in the rule set. Calculate the path confidence value and add a probability regularization term at each inference expansion. Terminate the expansion when the inference depth reaches the maximum length of the inference path or the path confidence value is lower than the preset threshold. S55. Calculate the normalized confidence score for the terminal nodes of all inference paths and output the root cause candidates sorted by confidence score.

7. The intelligent monitoring method for photovoltaic power plants based on digital twins according to claim 6, characterized in that, S6 specifically includes: S61. Input the root cause candidates into the digital twin simulation environment, and call the physical model of the corresponding device to generate simulation results; S62. Compare the difference between the simulated output power and the measured power, and select the root cause candidate with the smallest error as the final root cause determination result based on the simulation error. S63. Execute the corresponding maintenance actions based on the final root cause determination results, and synchronize the updated equipment parameters to the digital twin model.

8. A method for intelligent monitoring of photovoltaic power plants based on digital twins according to claim 7, characterized in that, Specifically, S7 includes: S71. After completing the operation and maintenance actions, obtain the updated equipment operation data and physical parameter sequence, compare the measured data with the simulation output of the digital twin model, and calculate the model error index. S72. Perform joint updates on the parameters of the DeepAR student model and the fuzzy Markov logic network based on the model error index.

Images