Photovoltaic power station remote monitoring system based on Internet of Things

Abstract

The invention relates to the technical field of photovoltaic power station remote monitoring, and discloses a photovoltaic power station remote monitoring system based on the Internet of Things, and the system comprises the steps: carrying out the importance evaluation of collected original data through a data value evaluation model, and generating a data importance weight; performing hierarchical compression processing on the data based on the data importance weight; sensitive data are processed by adopting an on-loop learning and error fully homomorphic encryption algorithm; original data features are protected through a differential privacy enhanced model updating mechanism; converting the operation of the machine learning model into polynomial operation supported by homomorphic encryption; a distributed data recovery mechanism is established at the cloud end, and multi-power-station collaborative analysis is supported in combination with a secure multi-party computing protocol. The technical problem that the existing method is difficult to balance the relationship among the data transmission efficiency, the data security and the analysis performance is solved, and efficient data transmission, full-life-cycle data security and multi-power-station security collaboration are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of new energy monitoring systems, and more specifically, to a method and system for remote monitoring of photovoltaic power stations based on the Internet of Things. Background Art

[0002] With the continuous expansion of the utilization of renewable energy, as an important part of clean energy, the installed capacity and the number of power stations of photovoltaic power generation are rapidly increasing globally. The safe and stable operation of photovoltaic power stations requires a perfect remote monitoring system to realize real-time monitoring, fault diagnosis and optimization adjustment of power generation equipment and system performance.

[0003] The existing remote monitoring technologies for photovoltaic power stations are mainly based on traditional SCADA systems and cloud computing architectures. The operation data of the power station is collected regularly and uploaded to the cloud server for processing and analysis. Related technologies include data acquisition methods with fixed sampling rates, traditional encryption transmission technologies, centralized data processing architectures, etc. Although these technologies can meet the basic monitoring requirements, they have obvious deficiencies in the face of new and complex scenarios.

[0004] However, the existing technologies face the following main problems: First, the monitoring data generated by photovoltaic power stations has the characteristics of high-dimensionality, heterogeneity, and multi-scale. The value density of different types of data varies significantly. The existing transmission methods adopt a unified sampling rate and compression strategy, and cannot dynamically adjust according to the importance of the data, resulting in difficult to guarantee the transmission quality of key data under limited bandwidth conditions; Second, the operation data of the power station includes commercially sensitive information such as electricity prices, power generation, and equipment parameters. Although the existing encryption methods can protect the transmission security, they usually need to be decrypted in the cloud and then analyzed and processed, making the sensitive data exposed during the analysis process, and there is a risk of data leakage; Third, in the scenario of collaborative analysis of multiple power stations, the existing distributed computing methods mainly rely on the central server to aggregate the original data or intermediate results of each node. The gradient information in the model aggregation process may be used to reverse the original data, and the communication overhead is large and the calculation delay is high, which is difficult to meet the requirements of real-time monitoring.

[0005] Therefore, there is an urgent need for a new method for remote monitoring of photovoltaic power stations that can balance the relationship between data transmission efficiency, data security, and analysis performance, so as to improve the overall performance and security of the monitoring system. Summary of the Invention

[0006] The present invention provides a method and system for remote monitoring of photovoltaic power stations based on the Internet of Things, which solves the technical problem that it is difficult to balance the relationship between data transmission efficiency, data security, and analysis performance in the existing technology.

[0007] The present invention discloses a method for remote monitoring of a photovoltaic power station based on the Internet of Things, including: evaluating the importance of the collected original data of the photovoltaic power station according to a data value evaluation model and generating a data importance weight, where the data importance weight is calculated based on the data change rate, the degree of abnormality, the prediction deviation, and the business relevance; performing hierarchical compression processing on the data based on the data importance weight, using low-loss compression for high-value data and high-compression ratio encoding for conventional data; processing sensitive data using the ring learning and error fully homomorphic encryption algorithm, and applying different strength encryption schemes to data with different sensitive levels in combination with a hierarchical encryption strategy; storing the power station data locally, protecting the original data features through a differentially private enhanced model update mechanism, and adding calibrated noise before transmitting the model parameters; converting the operations of the machine learning model into polynomial operations supported by homomorphic encryption, implementing the activation function using polynomial approximation, and reducing the encryption analysis delay through batch homomorphic operations and GPU parallel computing; establishing a distributed data recovery mechanism in the cloud, using statistical models and machine learning methods to recover detailed information from the highly compressed data, and supporting joint analysis of multiple power stations while protecting data privacy in combination with a secure multi-party computing protocol.

[0008] Further, in the step of evaluating the importance of the collected original data of the photovoltaic power station according to the data value evaluation model, the calculation formula of the data importance weight is: W(x) = α·Rc(x) + β·A(x) + γ·Dp(x) + δ·B(x), where Rc(x) represents the data change rate, A(x) represents the degree of abnormality, Dp(x) represents the prediction deviation, B(x) represents the business relevance, and α, β, γ, and δ are weight coefficients, satisfying α + β + γ + δ = 1.

[0009] Further, in the step of processing sensitive data using the ring learning and error fully homomorphic encryption algorithm, different encryption parameters are set for data with different sensitive levels: for data with a high sensitive level, the degree of the polynomial ring is set to a higher value, and the standard deviation of the error distribution is set to a larger value; for data with a medium sensitive level, the degree of the polynomial ring is set to a medium value, and the standard deviation of the error distribution is set to a medium value; for data with a low sensitive level, the degree of the polynomial ring is set to a lower value, and the standard deviation of the error distribution is set to a smaller value.

[0010] Further, in the step of protecting the original data features through a differentially private enhanced model update mechanism, it includes: training a machine learning model using the power station data on a local edge device to obtain the original model parameters; determining the noise intensity based on the privacy budget and data sensitivity, and adding Gaussian noise to the original model parameters; transmitting the model parameters with added noise to a coordination server for aggregation; and transmitting the aggregated model parameters back to each power station for updating the local model.

[0011] Further, in the step of converting the operations of the machine learning model into polynomial operations supported by homomorphic encryption, it includes: analyzing the basic operation types in the neural network model, mapping linear operations to addition and multiplication operations supported by homomorphic encryption; performing polynomial approximation on the non-linear activation functions in the neural network; utilizing the single instruction multiple data feature of fully homomorphic encryption to pack multiple plaintext data in a single ciphertext to achieve parallel processing; and through computational graph optimization, merging consecutive linear operations to reduce the generation and storage of intermediate results.

[0012] Further, in the step of establishing a distributed data recovery mechanism in the cloud, it includes: establishing a mapping relationship model between the original data and the compressed data; based on data correlation and time continuity, implementing the completion of missing data; using a deep learning model to learn the internal structure and features of the data to recover high-quality data from the compressed data; and integrating the output results of multiple recovery models to improve the recovery accuracy through an ensemble learning method.

[0013] Further, in the step of combining a secure multi-party computation protocol to support joint analysis among multiple power stations while protecting data privacy, it includes: constructing a secure multi-party computation protocol framework based on homomorphic encryption and zero-knowledge proof technologies; implementing analysis functions such as secure aggregation statistics, secure anomaly detection, and secure performance comparison; optimizing the protocol execution efficiency using a distributed computing framework; and establishing a joint knowledge base to store the general knowledge extracted from the data analysis of multiple power stations.

[0014] The present invention also discloses an Internet of Things-based remote monitoring system for photovoltaic power stations, including a data preprocessing module, a security encryption module, a distributed collaboration module, a security computing module, and a data recovery module, and each module is respectively used to implement the corresponding steps of the above method.

[0015] The Internet of Things-based remote monitoring method and system provided by the present invention have the following beneficial effects:

[0016] Significantly improved data transmission efficiency: By evaluating the value of the original data of the photovoltaic power station, implementing hierarchical compression processing based on the data importance weight, using low-loss compression for high-value data and high-compression ratio encoding for conventional data, and at the same time utilizing time series prediction technology to only transmit the prediction error. This method achieves a data compression rate of more than 90%, while maintaining 100% lossless detection of key events, reducing the bandwidth requirement by 85% and the storage cost by 70%.

[0017] Comprehensive privacy and security protection: Using the ring learning with errors (RLWE) fully homomorphic encryption algorithm to process sensitive data, and protecting the original data features through a differentially private enhanced model update mechanism to ensure the analysis and collaborative calculation of sensitive data without revealing the original data. This method guarantees the security of the entire data processing life cycle and meets the data security compliance requirements of the power industry.

[0018] Significantly enhanced analysis performance: By converting the operations of the machine learning model into polynomial operations supported by homomorphic encryption, implementing the activation function using polynomial approximation, and combining batch homomorphic operations and GPU parallel computing, the encryption calculation latency is significantly reduced. Experiments show that this method reduces the analysis processing time by 60%, supports near-real-time data analysis, and improves the efficiency of power plant operation and maintenance decision-making.

[0019] Improved multi-station collaboration ability: Based on the secure multi-party computing protocol, it supports multiple photovoltaic power plants to conduct joint analysis while protecting their respective data privacy, breaking through the limitations of data islands. Practical applications show that this method improves the anomaly detection accuracy by 25% and increases the fault warning lead time by 48 hours, significantly enhancing the operation level of the power plant group.

[0020] Enhanced system adaptability: Automatically optimizes the processing strategy according to network bandwidth, storage capacity, and application requirements, adapts to photovoltaic power plants of different scales and deployment environments, improves the system scalability and compatibility by 40%, and can meet the application scenario requirements of various types of photovoltaic power plants from small distributed ones to large-scale ground power plants. Description of the Drawings

[0021] Figure 1 is the overall flowchart of the remote monitoring method for photovoltaic power plants based on the Internet of Things of the present invention;

[0022] Figure 2 is the sub-step flowchart of the data preprocessing layer of the present invention;

[0023] Figure 3 is the sub-step flowchart of the secure encryption layer of the present invention;

[0024] Figure 4 is the sub-step flowchart of the distributed collaboration layer of the present invention;

[0025] Figure 5 is the sub-step flowchart of the secure computing layer of the present invention;

[0026] Figure 6 is the sub-step flowchart of the data recovery layer of the present invention. Detailed Embodiments

[0027] Overview of the solution of this application;

[0028] This application provides a remote monitoring method for photovoltaic power plants based on the Internet of Things, which effectively solves the above technical problems by establishing a multi-level data processing framework. It should be understood that according to the embodiments of this application, this method includes five key levels: data preprocessing layer, secure encryption layer, distributed collaboration layer, secure computing layer, and data recovery layer.

[0029] In the data preprocessing layer, according to the embodiments of the present application, this method is deployed on the edge devices of the power station Internet of Things, evaluates the value of the collected raw data, calculates the data importance weights in real time according to indicators such as data change rate, anomaly degree, prediction deviation, etc., processes high-value data using a low-loss compression algorithm, encodes conventional data with a high compression ratio, and at the same time uses local historical data to predict short-term trends and only transmits the prediction errors, thus significantly reducing the data transmission volume.

[0030] In the security encryption layer, according to the embodiments of the present application, this method uses the ring learning with errors (RLWE) fully homomorphic encryption algorithm to process sensitive data, supports direct addition and multiplication operations on encrypted state data, combines a hierarchical encryption strategy, and applies encryption schemes with different strengths according to the data sensitivity level, thereby balancing security and computational efficiency.

[0031] In the distributed collaboration layer, according to the embodiments of the present application, this method stores the power station data locally and only participates in model training when needed, protects the original data features through a model update mechanism enhanced by differential privacy, and adds calibrated noise before transmitting the model parameters, thus effectively preventing parameter reverse derivation.

[0032] In the secure computing layer, according to the embodiments of the present application, this method implements neural network processing in the encrypted domain, converts the operations of the machine learning model into polynomial operations supported by homomorphic encryption, uses polynomial approximation to implement the activation function, and reduces the encryption analysis delay through batch homomorphic operations and GPU parallel computing, thus ensuring real-time performance.

[0033] In the data recovery layer, according to the embodiments of the present application, this method establishes a distributed data recovery mechanism in the cloud, uses statistical models and machine learning methods to recover detailed information from highly compressed data, combines a secure multi-party computing protocol, and supports multiple photovoltaic power stations to conduct joint analysis while protecting their respective data privacy, thus improving the overall decision-making quality.

[0034] Application scenario description;

[0035] According to the embodiments of the present application, the provided remote monitoring method for photovoltaic power stations based on the Internet of Things is applicable to the operation monitoring, fault diagnosis, and performance optimization of various types of photovoltaic power stations. The application scenario of the present application is described below:

[0036] It should be noted that the technical solution of this application is mainly applied to the remote monitoring systems of distributed photovoltaic power stations and large-scale ground power stations. In a typical application scenario, data acquisition devices (such as intelligent inverters, weather stations, power quality analyzers, etc.) and Internet of Things edge computing devices are deployed at the photovoltaic power station site. The data acquisition devices are responsible for collecting power station operation data, including but not limited to: environmental data (such as irradiance, temperature, humidity, wind speed, etc.), device operation status data (such as inverter working status, component temperature, input and output voltage and current, etc.), and power generation and grid interaction data (such as active power, reactive power, power quality parameters, etc.).

[0037] According to an embodiment of this application, the Internet of Things edge computing device is connected to the data acquisition device through a wired or wireless communication method, receives the original data, and executes the data preprocessing and security encryption functions of this application. The processed data is transmitted to the cloud data center through the Internet, and data recovery, analysis, and multi-power station collaborative calculation are realized in the data center. Therefore, power station operation and maintenance personnel can access the monitoring platform through a Web application or a mobile terminal to view the real-time operation status, performance analysis report, and fault warning information of the power station.

[0038] It should be understood that this application is particularly applicable to the following scenarios:

[0039] Photovoltaic power stations in remote areas with limited network conditions: Such power stations usually have limited and unstable network bandwidth. By adopting the adaptive data transmission strategy of this application, the transmission quality of key data can be guaranteed under limited bandwidth.

[0040] Commercial and industrial photovoltaic power stations: Such power stations usually have high requirements for data security and privacy protection. By adopting the fully homomorphic encryption and differential privacy protection technologies of this application, it is ensured that sensitive commercial data is not leaked during the analysis process.

[0041] Large-scale photovoltaic power station groups: In scenarios where multiple power stations are managed by the same operator or need to perform collaborative analysis, by adopting the secure multi-party computing protocol of this application, group intelligent analysis and optimization are realized on the premise of protecting the data privacy of each power station.

[0042] The steps included in the method of this embodiment;

[0043] According to an embodiment of this application, the provided Internet of Things-based remote monitoring method for photovoltaic power stations includes the following steps:

[0044] Step S1: Adaptive preprocessing of data value perception;

[0045] According to an embodiment of this application, step S1 processes the original data of the photovoltaic power station collected based on the adaptive preprocessing mechanism of data value perception, and specifically includes the following sub-steps:

[0046] Step S1-1: The data value evaluation model calculates the data importance weight;

[0047] On the Internet of Things edge device, the data value evaluation model is used to evaluate the importance of the collected raw data and generate the data importance weight. This model comprehensively considers the following factors to calculate the data importance:

[0048] Data change rate: Calculate the change rate between the current data point and the historical data. The greater the change rate, the higher the importance. The calculation formula for the data change rate is:

[0049]

[0050] where x t represents the data value at time t, and x t-1 represents the data value at time t-1. ∈ is a small positive number to prevent the denominator from being zero.

[0051] In some embodiments, different change rate thresholds can be set based on different data types. For example, for light intensity data, a relatively high threshold (such as 15%) can be set; while for the temperature data of key power station equipment such as inverters, a relatively low threshold (such as 5%) can be set to improve sensitivity.

[0052] Degree of abnormality: Calculate the abnormality score of the data point. The higher the degree of abnormality, the higher the importance. The calculation formula for the abnormality score is:

[0053]

[0054] where μ w and σ w represent the mean and standard deviation of the data within the time window w respectively.

[0055] Optionally, the anomaly detection can also be combined with the normal operating range of the photovoltaic power station equipment for judgment. For example, in practical applications, when the output current of the inverter exceeds 80% or is lower than 10% of the rated value, regardless of the statistical distribution, it will be determined as high-importance data.

[0056] Prediction deviation: Calculate the difference between the actual value and the predicted value. The greater the deviation, the higher the importance. The calculation formula for the prediction deviation is:

[0057]

[0058] where represents the data value at time t predicted based on historical data.

[0059] In the specific application scenarios of photovoltaic power station monitoring, different prediction models can be adopted for different data types. For example, for solar irradiance data with obvious periodic changes, a seasonal time series model including weather factors can be used; while for device parameters such as inverter efficiency, a regression model based on device status can be used.

[0060] Business relevance: The business importance factor B(x t ) preset according to the data type and application requirements. For example, parameters related to fault diagnosis have a higher business relevance.

[0061] In some embodiments, the business relevance can be dynamically adjusted. For example, when an abnormal event occurs in a certain area of the power station, the system can automatically increase the business relevance of all monitoring points in that area to capture more data details related to faults.

[0062] Taking the above factors into account, the calculation formula for the data importance weight is:

[0063] W(x t ) = α·R c (x t ) + β·A(x t ) + γ·D p (x t ) + δ·B(x t )

[0064] Where α, β, γ, and δ are weight coefficients, satisfying α + β + γ + δ = 1, and can be dynamically adjusted according to the specific application scenario.

[0065] Optionally, in the application of large-scale photovoltaic power station groups, the weight coefficients can be automatically optimized through machine learning methods based on historical data and fault records. For example, for photovoltaic power stations in high-altitude areas, the weight coefficient of environmental factors can be increased; while for power stations in coastal areas, the weight coefficient of indicators related to equipment corrosion can be increased.

[0066] Step S1-2: Hierarchical compression processing based on data importance;

[0067] According to the data importance weight W(x t ) calculated in step S1-1, perform hierarchical compression processing on the data:

[0068] For high-value data with an importance weight higher than the threshold T h , adopt a low-loss compression algorithm, such as lossless compression or high-precision lossy compression, to ensure data quality. The compressed data is represented as:

[0069] C h (x t ) = Comp h (xt )

[0070] Among them, Comp h represents a low-loss compression function.

[0071] For medium-value data with importance weights between the thresholds T m and T h a balanced compression algorithm is adopted to achieve a balance between the compression ratio and data quality:

[0072] C m (x t ) = Comp m (x t )

[0073] Among them, Comp m represents a balanced compression function.

[0074] For regular data with importance weights lower than the threshold T m high compression ratio encoding is adopted:

[0075] C l (x t ) = Comp l (x t )

[0076] Among them, Comp l represents a high compression ratio encoding function.

[0077] Step S1-3: The time series data prediction model calculates the prediction error;

[0078] The time series data prediction model is used to analyze local historical data, predict short-term trends, calculate the prediction error, and further reduce the amount of transmitted data:

[0079] A lightweight time series prediction model, such as a simplified version of the autoregressive integrated moving average (ARIMA) model or the long short-term memory (LSTM) network, is trained on the edge device to predict the short-term trends of the data.

[0080] For data sequences that meet the prediction conditions, only the prediction error values are transmitted, rather than the original data or the fully compressed data:

[0081]

[0082] Among them, E t represents the prediction error at time t.

[0083] At the receiving end, the original data is reconstructed through the same prediction model and the received error values:

[0084]

[0085] Among them, x t ′ represents the reconstructed data value.

[0086] Step S1-4: The adaptive transmission strategy dynamically adjusts the transmission parameters;

[0087] Based on the network condition and system resources, dynamically adjust the data transmission strategy:

[0088] Monitor the current network bandwidth, latency, and system resource conditions, and generate a network status evaluation index N s .

[0089] According to the network status evaluation index and the data importance weight, dynamically adjust the data transmission priority, sampling rate, and compression parameters:

[0090] When the network condition is good, appropriately reduce the compression ratio to improve the data quality;

[0091] When the network condition is limited, increase the transmission priority of high-value data and reduce or suspend the transmission of low-value data;

[0092] When the network is extremely limited, only transmit abnormal event and critical alarm data;

[0093] The output result of this step is the photovoltaic power station data stream after value-aware preprocessing, including hierarchically compressed data and prediction error data, which realizes a significant reduction in data volume compared to the original data while maintaining the quality and integrity of high-value data.

[0094] Step S2: Hierarchical security protection of fully homomorphic encryption;

[0095] Step S2 implements hierarchical security protection based on the ring learning with errors (RLWE) fully homomorphic encryption algorithm to protect the sensitive data to be transmitted, specifically including the following sub-steps:

[0096] Step S2-1: Classification and marking of data sensitivity levels;

[0097] According to the data type and business characteristics, classify and mark the sensitivity levels of the data:

[0098] Establish a data sensitivity evaluation standard, analyze the data types of the photovoltaic power station, such as power generation, electricity price information, detailed power station parameters, inverter status, etc., and evaluate their sensitivity and protection requirements.

[0099] Divide the data into high-sensitivity levels (such as commercial data like electricity price, revenue, etc.), medium-sensitivity levels (such as detailed power generation, real-time equipment status, etc.), and low-sensitivity levels (such as environmental parameters, basic operating status, etc.).

[0100] Add a sensitivity level mark S to the datalevel , as the basis for subsequent encryption strategy selection.

[0101] Step S2-2: Parameter generation of the ring learning with errors fully homomorphic encryption algorithm;

[0102] Generate adaptive RLWE fully homomorphic encryption parameters according to the data sensitivity level and security requirements:

[0103] For highly sensitive data, generate encryption parameters with higher security strength to ensure sufficient security:

[0104] Degree n of the polynomial ring h Set to a higher value (e.g., n h = 8192)

[0105] Standard deviation σ of the error distribution h Set to a larger value;

[0106] Modulus q h Select an appropriate value that meets the security requirements;

[0107] In some embodiments, for commercially extremely sensitive data such as electricity prices and revenues, a higher-strength parameter configuration can be selected. For example, the degree of the polynomial ring can be increased to 16384, and a more complex error distribution can be used to provide a security strength similar to that of post-quantum cryptography.

[0108] For moderately sensitive data, generate encryption parameters that balance security and efficiency:

[0109] Degree n of the polynomial ring m Set to a medium value (e.g., n m = 4096)

[0110] Standard deviation σ of the error distribution m Set to a medium value;

[0111] Modulus q m Select a value that balances security and computational efficiency;

[0112] Optionally, in the actual deployment of a photovoltaic power station monitoring system, the encryption parameters for moderately sensitive data (such as detailed power generation and operating parameters) can be dynamically adjusted according to network conditions. For example, increase the security parameters during public network transmission, and appropriately reduce the parameters in a private network environment to improve performance.

[0113] For low-sensitive data, generate encryption parameters that focus more on efficiency:

[0114] Degree n of the polynomial ring l Set to a lower value (e.g., n l = 2048)

[0115] Standard deviation σ of error distribution l Set to a smaller value;

[0116] Modulus q l Select a value that meets basic security and high efficiency;

[0117] In typical application scenarios of a photovoltaic power station, environmental data (such as irradiance, temperature, etc.) is usually classified as low-sensitivity level, and lightweight encryption parameters can be used. For example, when the resources of edge computing devices are limited, the degree of the polynomial ring can be further reduced to 1024, and at the same time, a smaller modulus can be selected to reduce the computational complexity and memory requirements.

[0118] Parameter generation takes into account the requirements of subsequent homomorphic operations to ensure that data of different sensitivity levels can still perform cross-level homomorphic calculations.

[0119] In some embodiments, a parameter adaptive adjustment algorithm can be implemented to dynamically balance security and efficiency by monitoring the computational resource utilization rate and security threat level. For example, when a potential security threat is detected, the system can automatically increase the encryption parameter strength; while when the computational resources are tight, the encryption strength of non-critical data can be temporarily reduced to ensure the overall stable operation of the system.

[0120] Step S2-3: Execution of RLWE fully homomorphic encryption;

[0121] Using the generated parameters, perform RLWE fully homomorphic encryption operations:

[0122] For highly sensitive data x h , perform fully homomorphic encryption:

[0123] c h = Enc RLWE (x h , params h )

[0124] where Enc RLWE is the RLWE encryption algorithm, and params h is a high-security level parameter set.

[0125] For medium-sensitive data x m , perform fully homomorphic encryption:

[0126] c m = Enc RLWE (x m , params m )

[0127] where params m is a medium-security level parameter set.

[0128] For low-sensitivity level data x l , perform fully homomorphic encryption:

[0129] c l = Enc RLWE (x l , params l )

[0130] where params l is a parameter set of low security level.

[0131] Step S2-4: Key management and distribution;

[0132] Implement a secure key management and distribution mechanism:

[0133] Adopt a distributed key generation protocol to generate encryption keys, avoiding a single point from holding the complete key:

[0134] The public key pk is used for data encryption and can be widely distributed;

[0135] The private key sk is used for data decryption, stored in fragments, and requires multi-party authorization to reconstruct;

[0136] The evaluation key evk is used to support homomorphic multiplication operations and is only distributed to authorized computing nodes;

[0137] Implement a key rotation mechanism to regularly update encryption keys and reduce the risk of long-term key leakage.

[0138] Establish role-based access control to ensure that only authorized systems and users can access corresponding levels of data and keys.

[0139] The output result of this step is data processed by fully homomorphic encryption. These data can still be computationally analyzed in the encrypted state, support data analysis and mining without decryption, and effectively protect the sensitive data security of the photovoltaic power station.

[0140] Step S3: Differential privacy-enhanced distributed collaboration;

[0141] Step S3 implements a differential privacy-enhanced distributed collaboration mechanism to protect the original data features and support multi-power station collaborative analysis at the same time, specifically including the following sub-steps:

[0142] Step S3-1: Local data storage and access control;

[0143] Implement secure local data storage and access control on the edge devices of the photovoltaic power station:

[0144] Store the data preprocessed in Step S1 and encrypted in Step S2 in the local storage system to ensure that the original power station data is not directly uploaded to the cloud.

[0145] Implement fine-grained data access control and restrict access based on user permissions and data sensitivity levels:

[0146] Set different levels of access permissions, such as read-only permissions, aggregation analysis permissions, full access permissions, etc.;

[0147] Audit and record access behaviors, including access time, visitor information, and accessed data types;

[0148] Establish an anomaly detection mechanism for data access to identify potential unauthorized access attempts;

[0149] Implement data life cycle management, automatically execute data retention and deletion policies, and ensure that expired data is properly handled.

[0150] Step S3-2: Differential privacy parameter calculation;

[0151] Calculate the parameters required for differential privacy protection and determine the noise addition strategy:

[0152] Based on data characteristics and application scenarios, determine the privacy budget ∈, which controls the balance between data protection intensity and analysis availability:

[0153] A smaller ∈ value provides stronger privacy protection but may reduce the accuracy of data analysis;

[0154] A larger ∈ value provides higher analysis accuracy but relatively weaker privacy protection;

[0155] Calculate the sensitivity S, which represents the maximum impact that a change in a single data sample can have on the query result:

[0156] S = max D,D′ ||f(D) - f(D′)||1

[0157] where D and D′ are adjacent data sets that differ by one data sample, and f is the query function.

[0158] Select an appropriate noise distribution according to the query type and application requirements:

[0159] For aggregation queries, use noise from the Laplace distribution: Lap(S / ∈)

[0160] For query results in vector form, use noise from the Gaussian distribution: N(0, S 2 σ 2 )), where σ is related to the privacy budget ∈ and δ;

[0161] Step S3-3: Differential privacy enhanced model update;

[0162] Implement a model update mechanism with enhanced differential privacy to protect sensitive information in model parameters:

[0163] Train a machine learning model using power station data on a local edge device to obtain the original model parameters w i 。

[0164] In some embodiments, a suitable local model can be selected according to the scale and data characteristics of the photovoltaic power station. For example, for a small power station, a lightweight model such as a random forest or a simplified neural network can be used; while for a large power station, a more complex deep learning model such as a long short-term memory network (LSTM) or a graph neural network (GNN) can be used to capture the complex correlation relationships between devices.

[0165] Add calibrated noise before transmitting the model parameters to generate model parameters with differential privacy protection:

[0166]

[0167] Among them, is the model parameter after adding noise, and N(0,S 2 σ 2 ) represents Gaussian noise with a mean of 0 and a variance of S 2 σ 2 .

[0168] Optionally, different strategies can be adopted in the noise addition process. For example, in a photovoltaic power station anomaly detection model, smaller noise can be added to the parameters of the key layers (such as the feature extraction layer), while larger noise can be added to the non-key layers, so as to maintain the sensitivity of the model to abnormal states while protecting privacy.

[0169] Transmit the model parameters after adding noise to the coordination server for model aggregation:

[0170]

[0171] Among them, m is the number of power stations participating in collaborative training, is the aggregated model parameter.

[0172] In practical applications, for different types of photovoltaic power stations, a weighted average aggregation method can be adopted. For example, different weights can be assigned according to the power station scale, data quality, and historical contribution degree:

[0173]

[0174] Among them, v i is the weight coefficient of the i-th power station, satisfying

[0175] The aggregated model parameters are sent back to each power station for updating the local model, avoiding the direct sharing and exchange of raw data.

[0176] In some embodiments, an incremental learning strategy can be implemented to enable the local model to integrate global knowledge while retaining local characteristics. For example, for photovoltaic power stations with special environmental conditions (such as high altitude, desert or coastal areas), after applying the global model parameters, a small amount of local data can be used for fine-tuning to better adapt to the operating conditions in specific scenarios.

[0177] Step S3-4: Adaptive privacy budget allocation;

[0178] Implement an adaptive privacy budget allocation strategy according to data importance and analysis requirements:

[0179] Set different privacy budgets ∈ for different types of data analysis operations j , forming a privacy budget pool ∈ pool = ∑ j ∈ j .

[0180] Dynamically adjust the privacy budget allocation for each operation according to the importance and frequency of the analysis operation:

[0181] Allocate more privacy budgets to critical analysis operations (such as fault diagnosis);

[0182] Allocate fewer privacy budgets to routine analysis operations (such as performance statistics);

[0183] Implement a tracking and control mechanism for privacy budget usage to ensure that the total privacy budget does not exceed the preset limit and prevent privacy leakage caused by excessive queries.

[0184] The output result of this step is a distributed collaborative analysis framework with differential privacy protection. This framework can support collaborative analysis and model training among multiple power stations while protecting the privacy of the original data of each power station, effectively balancing the relationship between data analysis value and privacy protection.

[0185] Step S4: Neural network processing in the encrypted domain;

[0186] Step S4 implements neural network processing in the encrypted domain, enabling the machine learning model to directly process encrypted data, specifically including the following sub-steps:

[0187] Step S4-1: Convert neural network operations to polynomial operations supported by homomorphic encryption;

[0188] Convert the basic operations of the neural network to polynomial operations supported by fully homomorphic encryption:

[0189] Analyze the basic operation types in the neural network model, including linear transformations (matrix multiplication, vector addition) and non-linear activation functions (such as ReLU, Sigmoid, etc.).

[0190] Linear operations are directly mapped to operations supported by homomorphic encryption:

[0191] Homomorphic addition: Enc(a) + Enc(b) = Enc(a + b)

[0192] Homomorphic scalar multiplication: c·Enc(a) = Enc(c·a), where c is a constant;

[0193] Homomorphic multiplication: Enc(a)·Enc(b) = Enc(a·b), which requires the use of the evaluation key; evk

[0194] For complex matrix operations, decompose them into basic addition and multiplication operations to adapt to the computational characteristics of homomorphic encryption.

[0195] Step S4-2: Polynomial approximation of non-linear activation functions;

[0196] Use polynomial functions to approximate non-linear activation functions in the neural network:

[0197] For commonly used activation functions (such as ReLU, Sigmoid, Tanh, etc.), find appropriate polynomial approximations through Taylor expansion or other polynomial fitting methods:

[0198]

[0199] where f(x) is the activation function, a i are the polynomial coefficients, and d is the degree of the polynomial.

[0200] Optimize the polynomial coefficients a i , and achieve a balance between accuracy and computational complexity:

[0201] For the ReLU function: ReLU(x) = max(0, x), the following polynomial approximation can be used: ReLU(x) ≈ a0 + a1x + a2x 2 + a3x 3

[0202] For the Sigmoid function: The following polynomial approximation can be used: σ(x) ≈ b0 + b1x + b2x 2 + b3x 3 + b4x 4

[0203] Evaluate the accuracy of the polynomial approximation within a specific input range, and adjust the order and coefficients of the polynomial according to application requirements.

[0204] Step S4-3: Optimization of batch homomorphic operations;

[0205] Implement batch homomorphic operations to improve the efficiency of encrypted computing:

[0206] Utilize the SIMD (Single Instruction Multiple Data) feature of fully homomorphic encryption to pack multiple plaintext data in a single ciphertext for parallel processing:

[0207] Pack multiple scalar values into a vector; v = [v1, v2,..., v n

[0208] Encrypt the entire vector to obtain a single ciphertext; c = Enc(v)

[0209] Perform homomorphic operations on the ciphertext c to achieve parallel processing of all packed data;

[0210] Optimize the data layout and packing strategy to reduce cycle operations and rotation operations in encrypted computing:

[0211] Organize data according to the computing requirements to reduce cross-slot communication;

[0212] Utilize ciphertext replication and repetition patterns to avoid unnecessary rotation operations;

[0213] Implement computational graph optimization, merge consecutive linear operations, reduce the generation and storage of intermediate results, and lower the computational complexity.

[0214] Step S4-4: GPU parallel computing acceleration;

[0215] Utilize the GPU parallel computing power to accelerate the processing of neural networks in the encrypted domain:

[0216] Map computationally intensive operations in homomorphic encryption operations, such as large integer multiplication, number theoretic transform (NTT), etc., to the GPU for execution:

[0217] Utilize the parallel processing cores of the GPU to execute multiple independent homomorphic operations simultaneously;

[0218] Use the shared memory and registers of the GPU to optimize the data access pattern and reduce memory latency;

[0219] Implement a CPU-GPU collaborative computing strategy and allocate tasks according to the operation characteristics:

[0220] Allocate complex mathematical operations (such as polynomial multiplication, NTT transform) to the GPU;

[0221] Keep the control logic and lightweight operations on the CPU for execution;

[0222] ​Optimize CPU-GPU data transfer to reduce transfer overhead;

[0223] Develop an optimized algorithm library for different GPU architectures, adjust the calculation strategy according to hardware characteristics, and maximize the utilization rate of computing resources.

[0224] The output result of this step is a neural network model that can directly process encrypted data. This model realizes the analysis and processing of encrypted state data through fully homomorphic encryption technology, can perform intelligent analysis without decryption, and at the same time, through batch processing homomorphic operations and GPU acceleration technology, significantly reduces the latency of encrypted computing and improves the real-time performance of the system.

[0225] Step S5: Distributed data recovery and secure multi-party computation;

[0226] Step S5 realizes distributed data recovery and secure multi-party computation, recovers detailed information from highly compressed data, and supports secure collaborative analysis of multiple power stations, including the following sub-steps:

[0227] Step S5-1: Data recovery based on statistical models;

[0228] Implement a data recovery mechanism based on statistical models on the cloud platform:

[0229] Establish a mapping relationship model between the original data and the compressed data, and train a data recovery model by analyzing historical data samples:

[0230] For high-value data with low-loss compression, use a simple decompression algorithm to recover;

[0231] For regular data encoded with a high compression ratio, use a statistical model for information enhancement and detail recovery;

[0232] Based on data correlation and time continuity, complete the missing data:

[0233] Infer the possible values of the missing points using the data at adjacent time points;

[0234] Use the relationship constraints between related data items to optimize the rationality of the recovery result;

[0235] Dynamically evaluate the quality of the recovered data, and add a credibility mark to the recovered data through uncertainty estimation:

[0236]

[0237] where C(x t ′) is the credibility of the recovered data x t ′, is the context-based predicted value, and var(x t ) is the historical variance of this data item.

[0238] Step S5-2: Machine learning enhanced data restoration;

[0239] Apply machine learning methods to further enhance the data restoration effect:

[0240] Use deep learning models (such as convolutional neural networks, generative adversarial networks, etc.) to learn the internal structure and features of the data:

[0241] Training data: High-quality original data and corresponding compressed / degraded data pairs;

[0242] Output: High-quality data recovered from the compressed / degraded data;

[0243] Build dedicated restoration models for different types of photovoltaic power station data:

[0244] Environmental data restoration model: Reconstruct the complete time series of environmental parameters such as irradiance and temperature;

[0245] Equipment status restoration model: Restore the detailed operating status of equipment such as inverters and components;

[0246] Power generation performance restoration model: Reconstruct the details of key performance indicators such as power generation and efficiency;

[0247] Integrate the output results of multiple models and improve the restoration accuracy through ensemble learning methods:

[0248]

[0249] Among them, is the final restoration result, M j is the j-th restoration model, w j is the corresponding weight, is the compressed data, and k is the number of models.

[0250] Step S5-3: Implementation of secure multi-party computation protocol;

[0251] Implement a secure multi-party computation protocol to support joint analysis among multiple power stations while protecting data privacy:

[0252] Based on homomorphic encryption and zero-knowledge proof technologies, build a secure multi-party computation protocol framework:

[0253] Each power station acts as a participant and holds its own private data;

[0254] The protocol ensures the correctness of the calculation results and does not disclose the original data of each party;

[0255] In some embodiments, suitable protocol variants can be selected for photovoltaic power plant groups of different scales. For example, for a small-scale power plant group (such as within 10 power plants), a multi-party computation protocol based on secret sharing can be adopted, which has a relatively low communication complexity; while for a large-scale power plant group (such as dozens or hundreds of power plants), a hierarchical protocol structure can be adopted. After grouping the power plants, intra-group calculations are first performed, and then the results between groups are aggregated, thereby reducing the overall computational complexity.

[0256] Secure multi-party computation version for implementing key analysis functions:

[0257] Secure aggregation statistics: Calculate statistics such as the mean and variance of multi-power plant data;

[0258] Secure anomaly detection: Identify abnormal patterns in multi-power plant data without exposing the original data;

[0259] Secure performance comparison: Compare the performance indicators of different power plants and support benchmark testing;

[0260] Optionally, in the practical application of a photovoltaic power plant group, more specialized secure computing functions can be implemented. For example, secure fault attribution analysis allows multiple power plants to jointly analyze the root causes of certain common faults without sharing their respective detailed fault data; secure equipment life prediction, based on the equipment operation data of multiple power plants, can establish a more accurate equipment aging model on the premise of protecting the business secrets of each power plant.

[0261] Optimize the protocol execution efficiency using a distributed computing framework:

[0262] Decompose complex computing tasks into sub-tasks that can be executed in parallel;

[0263] Optimize the communication mode to reduce the number of interactions and the amount of data transmission between participants;

[0264] In the specific implementation of a photovoltaic power plant remote monitoring system, different computing modes can be selected according to the characteristics and urgency of the computing tasks. For example, for non-urgent tasks such as regular performance monitoring, a batch processing mode can be adopted to collect data within a certain period of time and then perform unified calculations to reduce system overhead; while for tasks with high timeliness requirements such as emergency fault diagnosis, a real-time computing mode can be enabled to ensure the timeliness of the analysis results even if it increases the consumption of computing resources.

[0265] Step S5-4: Joint decision-making and knowledge sharing;

[0266] Implement joint decision-making and knowledge sharing among multiple power plants on the premise of protecting data privacy:

[0267] Establish a joint knowledge base to store general knowledge extracted from multi-power plant data analysis, such as:

[0268] Common fault modes and their characteristics;

[0269] Typical modes of equipment performance degradation;

[0270] The influence law of environmental factors on power generation efficiency;

[0271] Implement a decision support system based on combined knowledge:

[0272] A prediction model trained based on the historical data of multiple power stations to provide more accurate power generation predictions;

[0273] Based on the fault case library of multiple power stations, improve the accuracy and efficiency of fault diagnosis;

[0274] Through comparative analysis, provide data support for the operation and maintenance optimization of power stations;

[0275] Implement a knowledge contribution and value distribution mechanism to encourage power stations to participate in data sharing:

[0276] Record the contribution degree of each power station to the combined knowledge base;

[0277] Allocate the permission to use advanced analysis functions according to the contribution degree;

[0278] Ensure the fairness and sustainability of knowledge sharing;

[0279] In a distributed photovoltaic power generation demonstration area in a coastal province, the Internet of Things-based remote monitoring method for photovoltaic power stations of the present invention is deployed in a power station group composed of 12 photovoltaic power stations of different scales. These power stations are distributed in areas such as City A, City B, City C, and City D, with a total installed capacity reaching 85 MW and an annual power generation of approximately 950 million kWh. The power station group includes 8 rooftop distributed power stations (single-station capacity 3 - 5 MW), 3 industrial and commercial building integrated power stations (single-station capacity 5 - 10 MW), and 1 small ground power station (capacity 15 MW).

[0280] An edge computing gateway device is deployed in each power station, equipped with a dual-core ARM processor, 4 GB of memory, and 128 GB of storage space, and is connected to the cloud platform through a 4G / 5G network. The on-site data acquisition system of each power station collects data every 5 seconds, including environmental data (irradiance, temperature, humidity, etc.), inverter operation parameters (input / output voltage and current, power, conversion efficiency, etc.), busbar box parameters, grid parameters, and other equipment status data. Each power station generates approximately 500 MB of raw data per day.

[0281] This area is characterized by rich solar resources but cloudy, rainy and overcast weather along the coast. The power stations are relatively scattered, and some are located in suburbs with limited network conditions. At the same time, as a group of commercially operated PV power stations, there are high requirements for the data security, real-time analysis and operation and maintenance efficiency of the system. The operator hopes to solve the three main problems currently faced through new technologies: First, the low data transmission efficiency under limited network conditions; second, the protection of commercially sensitive information during data sharing among multiple power stations; third, the computational efficiency and real-time challenges during collaborative analysis.

[0282] Implementation example of data preprocessing layer

[0283] An adaptive preprocessing system for data value perception was deployed on the edge computing gateway of a 5MW PV power station on the roof of a commercial building in City A. The system processes the data of 75 data acquisition points in real time, including 20 environmental monitoring points, 40 inverter monitoring points and 15 grid parameter monitoring points.

[0284] In the practical application of the data value evaluation model, the weight coefficients of each parameter were determined according to historical operation experience: the weight of data change rate α = 0.35, the weight of abnormality degree β = 0.4, the weight of prediction deviation γ = 0.15, and the weight of business relevance δ = 0.1. The system sets different basic values of business relevance for various data parameters: the parameters related to inverter faults (such as harmonic content, IGBT temperature) are set to 0.8 - 0.9, the parameters related to power generation are set to 0.6 - 0.7, and the environmental parameters are set to 0.3 - 0.5.

[0285] During actual operation, when the output current of an inverter in a certain area suddenly changes (jumping from 45% to 85% of the rated value), the system quickly calculates its importance weight: the data change rate Rc = 0.89 (large change amplitude), the abnormality degree A = 2.3 (exceeding the normal range), the prediction deviation Dp = 0.76 (large deviation from the predicted value), and the business relevance B = 0.85 (key equipment parameters). The comprehensive calculation gives the importance weight W of this data point as W = 0.35×0.89 + 0.4×2.3 + 0.15×0.76 + 0.1×0.85 = 1.361, exceeding the high-value data threshold Th = 1.0.

[0286] For such high-value data, the system automatically processes it using a lossless compression algorithm to retain full precision. At the same time, the system increases the sampling rate of all data points in the inverter area from 5 seconds per time to 1 second per time temporarily, and adjusts the transmission priority of the relevant data to the highest. At the same time, for the environmental data in the normal operation area (such as the irradiance is stable at 850W / m 2 , the change rate is less than 5%), the calculated importance weight is only 0.25, and the system automatically applies high compression ratio coding (20:1) and reduces the sampling frequency to 30 seconds per time.

[0287] During network bandwidth fluctuations (such as the network congestion period from 9:00 to 11:00 on weekdays), the system automatically adjusts the transmission strategy according to the real-time monitored network conditions: when the bandwidth drops to 40% of the original, the system increases the sampling interval of environmental data and normal operation device data to 60 seconds, and at the same time applies a higher compression ratio (35:1); only retains the original transmission frequency and quality of high-value data, reduces the total daily data transmission volume from the original plan of 500MB to 125MB, and at the same time ensures 100% complete transmission of key event data.

[0288] Implementation example of the security encryption layer

[0289] In a 10MW photovoltaic power station in an industrial park in City B, the ring learning with errors fully homomorphic encryption algorithm is applied to protect sensitive data. The operation data of this power station contains multiple levels of sensitive information such as electricity price information, power generation revenue, real-time power generation volume, and key equipment parameters. These data need to be uploaded to the cloud platform for analysis, but there is a risk of commercial secret leakage.

[0290] The system first classifies different data according to the sensitive level: marks the electricity price and revenue data (about 5% of the total data volume) as high sensitive level; marks the power generation volume, key equipment operation parameters, and performance indicators (about 35%) as medium sensitive level; marks the environmental data and basic operation status (about 60%) as low sensitive level.

[0291] For the electricity price data marked as high sensitive level, the system configures RLWE encryption parameters with high security strength: the degree n of the polynomial ring = 8192, the standard deviation σ of the error distribution = 3.2, the modulus q = 2^40 - 2^20 + 1, and the security level can resist quantum computing attacks. For the power generation volume data of medium sensitive level, medium security strength parameters are configured: the degree n of the polynomial ring = 4096, the standard deviation σ of the error distribution = 2.5, the modulus q = 2^30 - 2^10 + 1. The low sensitive level environmental data uses lightweight parameters: the degree n of the polynomial ring = 2048, the standard deviation σ of the error distribution = 1.8, the modulus q = 2^20 - 2^10 + 1.

[0292] In actual operation, the system processes about 150,000 encrypted data records every day. Although the high-level encrypted data only accounts for 5%, the encryption calculation overhead accounts for 30% of the total overhead. To reduce the calculation overhead, the system implements a dynamic parameter adjustment strategy: during the non-power generation period at night, the medium sensitive level data is automatically downgraded to use lightweight parameters for encryption; while during the period with large electricity price fluctuations (such as the peak-valley electricity price switching period), the encryption intensity of relevant data is temporarily increased.

[0293] In terms of key management, the system adopts a distributed key generation and storage scheme. The public key is publicly shared among nodes, while the private key is divided into five parts and stored in three management terminals with different permission levels respectively. The "3-out-of-5" threshold scheme is adopted, and at least three key fragments are required to recover the complete private key. The system also sets up an automatic key rotation mechanism to update all encryption keys every 30 days and retains historical keys to ensure the decryptability of historical data.

[0294] Implementation example of the distributed collaboration layer

[0295] A distributed collaboration mechanism enhanced by differential privacy was implemented among five distributed photovoltaic power stations in City C and City D. The installed capacities of these five power stations are 4.2MW, 3.8MW, 5.5MW, 6.7MW, and 4.5MW respectively. They are managed by the same operator but have different deployment environments and need to jointly participate in the training of the fault prediction model without sharing the original data.

[0296] The system deploys a lightweight fault prediction model on the local edge devices of each power station. It adopts a convolutional neural network structure, including four convolutional layers and two fully connected layers. The total number of model parameters is about 1.2 million. Each power station uses the local two-month historical data (about 12GB data volume) to train the local model and extracts key features such as abnormal inverter temperature, output power fluctuation, and current harmonic change.

[0297] In terms of differential privacy protection, the system sets reasonable privacy budgets ε = 0.8 and δ = 10^-5 according to the data sensitivity and adds noise using the Gaussian mechanism before uploading the model parameters. Taking the inverter temperature prediction model as an example, its sensitivity S is calculated to be about 0.23, and the system adds Gaussian noise with a standard deviation of σ = 0.36 to the model weight parameters. To optimize the performance, the system adopts a differential noise strategy for different layer parameters: adding less noise (σ = 0.28) to the key feature extraction layer (the first two convolutional layers) and adding more noise (σ = 0.42) to the subsequent layers.

[0298] The coordination server aggregates the model parameters of each power station by weighted average and assigns weights according to the data quality and scale of each power station: the power station with an installed capacity of 6.7MW has a weight of 0.28, and the weights of other power stations are 0.19, 0.17, 0.21, and 0.15 respectively. The aggregated model parameters are sent back to each power station through a secure channel for updating the local model. The aggregation process is carried out once a week. The single aggregation process takes about 4.5 minutes, and the transmitted parameter data volume is about 15MB.

[0299] To further protect privacy, the system implements an adaptive privacy budget allocation strategy. For the key analysis task of inverter fault diagnosis, 50% of the privacy budget is allocated; for the analysis task of power generation efficiency optimization, 30% of the budget is allocated; for the conventional performance monitoring task, only 20% of the budget is allocated. The system also establishes a privacy budget usage tracking mechanism. When the budget consumption of a certain type of analysis reaches the upper limit, the parameter accuracy of this type of analysis is automatically reduced or the relevant query is suspended.

[0300] Implementation example of the secure computing layer

[0301] An encrypted domain neural network processing system is implemented on the cloud platform of the power station group, enabling the machine learning model to directly analyze data in the encrypted state without the decryption process, thus eliminating potential security risks in the data analysis process. This system is mainly applied to the 5MW industrial and commercial rooftop power station in Area A, processing encrypted power generation data and business-sensitive information from this power station.

[0302] The system converts the deep neural network models for power station performance evaluation and anomaly detection into forms that support homomorphic encryption calculations. Taking the inverter anomaly detection model as an example, the original model contains 3 fully connected layers and ReLU activation functions. The input is a 20-dimensional device parameter vector, and the output is an anomaly probability score. The system first directly maps the linear layers in the model to matrix multiplication and vector addition operations supported by homomorphic encryption without changing the network structure.

[0303] For the non-linear ReLU activation function, the system uses a 7th-order polynomial approximation: ReLU(x)≈0.47+0.5x+0.09x 2 -0.02x 3 +0.003x 4 -0.0004x 5 +0.00002x 6 . In practical applications, the average error of this approximation in the interval [-4,4] is less than 0.06, meeting the accuracy requirements for anomaly detection. The system also optimizes the polynomial coefficients for different types of data: uses a 5th-order polynomial (average error 0.08) for temperature data to speed up calculations; uses an 8th-order polynomial (average error 0.04) for current data to improve accuracy.

[0304] In terms of batch homomorphic operations, the system utilizes the SIMD feature to simultaneously process 16 data points in a single ciphertext. In the specific implementation, the same type of parameters (such as output voltage) of 16 inverters are packed into a vector and then encrypted into a single ciphertext. Performing the same operation on each ciphertext enables parallel processing, reducing the operation time from 224 milliseconds in sequential processing to 26 milliseconds, with an acceleration ratio of 8.6 times.

[0305] To further improve performance, the system implements parallel computing acceleration on the NVIDIA V100 GPUs of cloud servers. The large integer polynomial multiplication operation in homomorphic encryption is mapped to the GPUs for execution, making full use of its 3,584 CUDA cores to process multiple independent polynomial multiplications simultaneously. In actual operation, GPU acceleration reduces the inference latency of the encrypted domain neural network from the original 1,200 milliseconds to 320 milliseconds, meeting the requirements of quasi-real-time monitoring (latency less than 500 milliseconds).

[0306] The system also optimizes the computational graph, analyzes the neural network model structure, and merges consecutive linear operations to reduce the number of multiplications. For example, converting "linear layer 1 → linear layer 2" into a single equivalent linear transformation simplifies the operation that originally required two matrix multiplications into one, reducing the intermediate result storage requirement from the original 42 MB to 6 MB and reducing the memory occupancy on edge devices by 85%.

[0307] Corresponding to the above-mentioned method for remote monitoring of photovoltaic power stations based on the Internet of Things, the present application also provides a remote monitoring system for photovoltaic power stations based on the Internet of Things, including:

[0308] A data preprocessing module, which is used to be deployed on the edge devices of the power station Internet of Things, evaluate the value of the collected raw data, calculate the data importance in real time according to indicators such as data change rate, anomaly degree, prediction deviation, etc., perform low-loss compression on high-value data, use high compression ratio encoding for conventional data, and at the same time predict short-term trends using local historical data and only transmit the prediction error, significantly reducing the data transmission volume.

[0309] A security encryption module, which is used to process sensitive data using the ring learning with errors (RLWE) fully homomorphic encryption algorithm, support direct addition and multiplication operations on encrypted state data, and combine a hierarchical encryption strategy to apply different strength encryption schemes according to the data sensitivity level to balance security and computational efficiency.

[0310] A distributed collaboration module, which is used to store the power station data locally and only participate in model training when needed, protect the original data features through a model update mechanism enhanced by differential privacy, and add calibrated noise before transmitting the model parameters, expressed as: where w i is the local model parameter, S is the sensitivity, and σ is the privacy budget.

[0311] A security computing module, which is used to implement encrypted domain neural network processing, convert the operations of the machine learning model into polynomial operations supported by homomorphic encryption, and use polynomial approximation to implement the activation function: And reduce the encryption analysis latency through batch homomorphic operations and GPU parallel computing to ensure real-time performance.

[0312] A data recovery module is used to establish a distributed data recovery mechanism in the cloud, recover detailed information from highly compressed data by using statistical models and machine learning methods, and support joint analysis by multiple power stations while protecting data privacy in combination with a secure multi-party computation protocol.

[0313] The working processes between the modules correspond to the steps described in the foregoing method embodiments and will not be elaborated herein.

Claims

1. A photovoltaic power station remote monitoring method based on the Internet of Things, characterized in that: The following steps are involved: According to the data value assessment model, the importance of the collected photovoltaic power station raw data is assessed and the data importance weight is generated. The data importance weight is calculated based on the data change rate, abnormality, prediction deviation and business relevance; Based on the data importance weight, the data is compressed in a hierarchical manner, low-loss compression is used for high-value data, and high-compression ratio encoding is used for regular data; The ring learning and error-free fully homomorphic encryption algorithm is used to process sensitive data, and encryption schemes of different strengths are applied to data of different sensitivity levels in combination with the layered encryption strategy; The power plant data is stored locally, the original data features are protected through a differential privacy-enhanced model update mechanism, and calibrated noise is added before the model parameters are transmitted; Convert the operations of the machine learning model into polynomial operations supported by homomorphic encryption, use polynomial approximation to implement the activation function, and reduce the encryption analysis latency through batch homomorphic operations and GPU parallel computing; A distributed data recovery mechanism is established in the cloud, and statistical models and machine learning methods are used to recover detailed information from highly compressed data. The secure multi-party computing protocol is combined to support joint analysis among multiple power stations while protecting data privacy.

2. The method for remote monitoring of a photovoltaic power station based on the Internet of Things according to claim 1, characterized in that: In the step of evaluating the importance of the collected PV power station raw data according to the data value evaluation model, the calculation formula of the data importance weight is: W(x)=α·Rc(x)+β·A(x)+γ·Dp(x)+δ·B(x) Among them, Rc(x) represents the data change rate, A(x) represents the degree of abnormality, Dp(x) represents the prediction deviation, B(x) represents the business relevance, α, β, γ and δ are weight coefficients, satisfying α+β+γ+δ=1.

3. The photovoltaic power station remote monitoring method based on the Internet of Things according to claim 1 is characterized in that: In the step of processing sensitive data using the ring learning and error-free fully homomorphic encryption algorithm, different encryption parameters are set for data of different sensitivity levels: For data with high sensitivity levels, the degree of the polynomial ring is set to a higher value and the standard deviation of the error distribution is set to a larger value; For data with medium sensitivity level, the degree of the polynomial ring is set to medium value and the standard deviation of the error distribution is set to medium value; For data with low sensitivity levels, the degree of the polynomial ring is set to a lower value and the standard deviation of the error distribution is set to a smaller value.

4. The photovoltaic power station remote monitoring method based on the Internet of Things according to claim 1 is characterized in that: The steps to protect the original data features through the model update mechanism enhanced by differential privacy include: Use power plant data to train machine learning models on local edge devices to obtain original model parameters; Determine the noise intensity based on the privacy budget and data sensitivity, and add Gaussian noise to the original model parameters; The model parameters after adding noise are transmitted to the coordination server for aggregation; The aggregated model parameters are transmitted back to each power station for updating the local model.

5. The photovoltaic power station remote monitoring method based on the Internet of Things according to claim 1 is characterized in that: The steps to convert the operations of the machine learning model into polynomial operations supported by homomorphic encryption include: Analyze the basic operation types in the neural network model and map linear operations to addition and multiplication operations supported by homomorphic encryption; Polynomial approximation of nonlinear activation functions in neural networks; Utilize the single instruction multiple data feature of fully homomorphic encryption to pack multiple plaintext data into a single ciphertext to achieve parallel processing; By optimizing the computational graph, continuous linear operations are merged to reduce the generation and storage of intermediate results.

6. The photovoltaic power station remote monitoring method based on the Internet of Things according to claim 1 is characterized in that: The steps to establish a distributed data recovery mechanism in the cloud include: Establish a mapping relationship model between original data and compressed data; Based on data correlation and time continuity, missing data can be completed; Use deep learning models to learn the intrinsic structure and features of data and recover high-quality data from compressed data; The output results of multiple recovery models are combined to improve the recovery accuracy through ensemble learning methods.

7. The method for remote monitoring of photovoltaic power stations based on the Internet of Things according to claim 1, characterized in that: The steps of combining the secure multi-party computing protocol to support joint analysis by multiple power stations while protecting data privacy include: Based on homomorphic encryption and zero-knowledge proof technology, a secure multi-party computing protocol framework is built; Realize security aggregation statistics, security anomaly detection and security performance comparison and analysis functions; Use a distributed computing framework to optimize protocol execution efficiency; Establish a joint knowledge base to store common knowledge extracted from multi-plant data analysis.

8. The method for remote monitoring of photovoltaic power stations based on the Internet of Things according to claim 1, characterized in that: Also includes: In the preprocessing layer, the time series data prediction model is used to analyze local historical data, predict short-term trends, calculate prediction errors, and only transmit prediction error values ​​instead of original data; Dynamically adjust data transmission strategies based on network conditions and system resources, increase the transmission priority of high-value data when network conditions are limited, and reduce or suspend the transmission of low-value data.

9. The photovoltaic power station remote monitoring method based on the Internet of Things according to claim 1, characterized in that: Also includes: Implement fine-grained data access control to restrict access based on user permissions and data sensitivity levels; Audit access behavior, including access time, visitor information, and access data type; Establish anomaly detection mechanisms for data access to identify potential unauthorized access attempts; Implement data lifecycle management to automate data retention and purge policies.

10. A photovoltaic power station remote monitoring system based on the Internet of Things, characterized in that: include: The data preprocessing module is used to evaluate the value of the collected raw data and perform hierarchical compression processing on the data according to the weight of data importance; The security encryption module is used to process sensitive data using the ring learning and error-free fully homomorphic encryption algorithm, and to apply encryption schemes of different strengths to data of different sensitivity levels; A distributed collaboration module that stores power plant data locally and protects original data features through a model update mechanism enhanced by differential privacy; The secure computing module is used to convert the operations of the machine learning model into polynomial operations supported by homomorphic encryption, and reduce the latency of encryption analysis through batch homomorphic operations and GPU parallel computing; The data recovery module is used to establish a distributed data recovery mechanism in the cloud, and combined with the secure multi-party computing protocol, it supports joint analysis by multiple power stations while protecting data privacy.

Images