Intelligent management method and system for operation and maintenance of energy station

Abstract

The invention discloses an intelligent management method and system for operation and maintenance of an energy station, and belongs to the technical field of energy station management, and the method comprises the steps: obtaining a physical operation rule of energy equipment; constructing a mechanism-constrained deep learning model, and embedding the physical operation rule of the energy equipment into a neural network training process, so that the model maintains high-reliability prediction in a small sample and noise data scene; determining an energy data type according to the physical operation rule of the energy equipment; aiming at different energy data types, a hierarchical feature extraction network is designed, key features are automatically identified and strengthened, and optimal combination of cross-energy features is realized through real-time weight adjustment; fusing the time domain trend, spatial domain distribution and frequency domain characteristics of the equipment operation data, and constructing a three-dimensional diagnosis matrix; and constructing an edge-cloud coevolution mechanism. The method has the effect of improving the operation and maintenance intelligence level and the decision-making efficiency.

Description

Technical Field

[0001] The present application relates to the technical field of energy station management, and in particular to an intelligent management method and system for energy station operation and maintenance. Background Art

[0002] Currently, energy stations are the core infrastructure for energy production and supply, encompassing thermal power plants, wind farms, photovoltaic power plants, nuclear power plants, energy storage plants, oil and gas fields, and transmission and distribution grids. With the global energy structure shifting toward cleaner, lower-carbon energy sources and the widespread application of intelligent technologies, the operation and maintenance of energy stations has become a critical component in ensuring energy security, improving efficiency, and reducing costs.

[0003] Energy site operations and maintenance (O&M) encompasses the full lifecycle management of renewable energy power generation facilities (wind, photovoltaic, nuclear, etc.) and energy storage equipment, encompassing monitoring, maintenance, troubleshooting, and technical upgrades. Core responsibilities include tracking equipment operating status to ensure power generation efficiency and safety, as well as regular inspections, component replacement, and system optimization to prevent unexpected failures.

[0004] However, new energy stations have diverse types of power generation facilities, and different power generation facilities require different monitoring data. Previous energy station operation and maintenance methods are difficult to effectively handle these complex, multi-source, and noisy data, resulting in low prediction accuracy and poor real-time performance of energy station operation and maintenance. Summary of the Invention

[0005] In order to improve the intelligence level of operation and maintenance and decision-making efficiency, the present application provides an intelligent management method and system for energy station operation and maintenance.

[0006] The present application provides an intelligent management method and system for energy station operation and maintenance, which adopts the following technical solutions:

[0007] In a first aspect, the present application provides an intelligent management method for energy station operation and maintenance, comprising the following steps:

[0008] Obtain the physical operating laws of energy equipment;

[0009] Build a mechanism-constrained deep learning model that embeds the physical operating laws of energy equipment into the neural network training process, enabling the model to maintain high-confidence predictions in scenarios with small samples and noisy data;

[0010] Determining the energy data type according to the physical operating laws of the energy equipment;

[0011] Design a hierarchical feature extraction network for different energy data types to automatically identify and enhance key features, and achieve the optimal combination of cross-energy features through real-time weight adjustment;

[0012] Integrate the time domain trend, spatial distribution and frequency domain characteristics of equipment operation data to build a three-dimensional diagnostic matrix;

[0013] Build an edge-cloud co-evolution mechanism, deploy lightweight feature extraction models at the edge, and process high-frequency data in real time; the cloud aggregates knowledge from multiple sites through transfer learning, dynamically updates the global model, and reversely optimizes the edge algorithm.

[0014] Furthermore, after the step of obtaining the physical operating laws of the energy device, the method further includes:

[0015] Perform fault logic mining based on the physical operation rules of the energy equipment, wherein the fault logic mining includes fault tree analysis and failure mode and effects analysis (HFMEA);

[0016] Constructing a health scoring system through scoring methods, including logistic regression model, random forest, and support vector machine;

[0017] Build a multi-dimensional health status model to comprehensively evaluate the health status of equipment;

[0018] According to the analysis results of the fault logic mining and health scoring system, the fault factor score value and the corresponding level are determined to achieve a multi-dimensional health assessment.

[0019] Furthermore, the step of obtaining the physical operating law of the energy device specifically includes:

[0020] Acquire meteorological satellite data and combine it with the CAMS atmospheric model to calculate irradiance parameters in real time, including horizontal irradiance (GHI), direct irradiance (DNI), and tilted irradiance (GTI);

[0021] An improved clear sky model is used to separate irradiance data under clear and cloudy conditions to improve short-term forecast accuracy;

[0022] Acquiring monitoring information, including terrain elevation data, historical meteorological data, and real-time sensor data;

[0023] Data cleaning and preliminary calibration are performed through the edge gateway to reduce redundant data transmission;

[0024] Build a dynamic virtual weather station and use Kalman filtering or Bayesian algorithm to optimize data deviation;

[0025] The feature importance is analyzed through random forest, the time series deviation is corrected through LSTM network, and the calibrated irradiation value is output;

[0026] Combining the gray-level co-occurrence matrix of satellite cloud images with the ARIMA model to predict cloud movement, the LSTM network is used to dynamically correct the numerical weather forecast results to achieve irradiance prediction;

[0027] Edge servers deploy lightweight prediction models, and edge gateways integrate data aggregation, anomaly detection, and fault warning functions to reduce the amount of uplink data.

[0028] Furthermore, after the step of designing a hierarchical feature extraction network for different energy data types to automatically identify and enhance key features, the method further includes:

[0029] Based on irradiation parameters and monitoring information, the inverter output strategy is dynamically adjusted, and the MPPT tracking efficiency is optimized through the PID algorithm to increase power generation;

[0030] Build an irradiation-energy production correlation model, input historical irradiation data, component attenuation parameters, and cleaning frequency, and output the equipment health index (HI) for life prediction and automatic generation of operation and maintenance reports and cleaning recommendations.

[0031] Furthermore, the step of automatically generating an operation and maintenance report and cleaning suggestions specifically includes:

[0032] Acquiring dust accumulation status information, wherein the dust accumulation status information includes infrared sensor data, laser scattering data, and image recognition data;

[0033] According to the dust accumulation status information, multimodal data is integrated to improve the assessment accuracy;

[0034] Use convolutional neural networks to train image recognition models and output dust accumulation coverage and power generation efficiency loss rate in real time;

[0035] Automatically adjust dust accumulation alarm threshold based on historical data and meteorological conditions;

[0036] Optimize the cleaning strategy with the goal of maximizing net profit. The constraints include:

[0037] Max Profit=∑(P clean |ΔE·T sun Price)-C clean ·N clean

[0038] Among them, MaxProfit represents the maximum net profit, P clean is the power improvement rate after cleaning, ΔE represents the increase in power generation after cleaning, T sun refers to the sunshine time, Price refers to the electricity price, C clean is the cost of each cleaning, N clean Refers to the number of cleanings;

[0039] Using the REINFORCE algorithm, input parameters such as weather forecast (rainfall probability), electricity price period, dust accumulation rate, etc., and output the optimal cleaning time window;

[0040] When the real-time calculated power generation loss cost is greater than or equal to the single cleaning cost, the cleaning task is automatically triggered.

[0041] Furthermore, the step of automatically triggering the cleaning task further includes:

[0042] The cleaning strategy is optimized through a reinforcement learning algorithm, taking the cleanliness-corrected daily intake (DNI) as input and incorporating cleaning costs into long-term maintenance costs. The optimal cleaning strategy is achieved through simulation and parameter optimization.

[0043] The particle swarm algorithm is used to determine the optimal cleaning time interval, thereby improving the net cleaning benefit;

[0044] Using the K-means clustering algorithm to analyze weather data from photovoltaic power plants, combined with the dust loss coefficient and economic threshold, the cleaning cycle is dynamically adjusted to balance power generation loss and cleaning costs.

[0045] Through linear programming models and iterative algorithms, the amount of detergent used and water consumption during the cleaning process are optimized to achieve the lowest cost and the best cleaning effect;

[0046] Utilizing big data analysis and AI technology, we evaluate cleaning effectiveness based on power generation data before and after cleaning, and continuously optimize cleaning strategies to improve economic benefits.

[0047] By calculating the cleaning cost and benefit ratio, select the cleaning plan with the highest benefit to cost ratio.

[0048] Furthermore, after the step of automatically triggering the cleaning task, the method further includes:

[0049] Adopting a tightly coupled solution of LiDAR, visual cameras, and inertial measurement units, it constructs a 3D point cloud map through front-end laser odometer and back-end map optimization technology.

[0050] Use laser SLAM equipment to collect point cloud data and combine it with visual texture mapping technology to generate a three-dimensional model with color information;

[0051] For dynamic environments, closed-loop detection algorithms and dynamic window methods are introduced to update maps in real time;

[0052] Classify the clean area into levels according to the complexity of the site terrain, including flat area, complex area, and dynamic area;

[0053] For the flat area, a rasterized map is used to mark the conventional cleaning path; for the complex area, an octree map is constructed to mark the obstacle collision risk through three-dimensional voxels; for the dynamic area, SLAM is used to perceive moving objects in real time to trigger dynamic map updates;

[0054] Combining Euclidean distance, Manhattan distance and height influence factor, we define the energy cost function:

[0055] f(n)=g(n)+λ1·h distance (n)+λ2·h height (n)

[0056] Among them, g(n) represents the actual energy consumption accumulated from the starting point to the current node n, h distance (n) represents the straight-line distance from the current node n to the end point, h height (n) represents the slope at node n, λ1 and λ2 are weight coefficients;

[0057] Adaptively adjust the search step size based on the complexity of the area. Use a large step size to accelerate the search in flat areas, and a small step size to avoid obstacles in complex areas.

[0058] Quasi-uniform B-spline curves are used to smooth the broken line path to reduce the number of turns and energy loss;

[0059] Combining spiral scanning with reciprocating paths, it achieves non-repetitive coverage through regional decomposition and reduces ineffective movement;

[0060] The cleaning order is dynamically adjusted based on the Markov decision process (MDP), giving priority to high-pollution areas and reducing the overall moving distance.

[0061] In a second aspect, the present application provides an intelligent management system for energy station operation and maintenance, including:

[0062] Physical operation law acquisition module, used to obtain the physical operation law of energy equipment;

[0063] The learning model construction module is used to build a deep learning model with mechanism constraints, embedding the physical operation laws of energy equipment into the neural network training process, so that the model can maintain high-confidence predictions in scenarios with small samples and noisy data;

[0064] An energy type determination module, configured to determine the type of energy data according to the physical operating rules of the energy equipment;

[0065] Key feature extraction module, which is used to design a hierarchical feature extraction network for different energy data types, automatically identify and enhance key features, and achieve the optimal combination of cross-energy features through real-time weight adjustment;

[0066] The diagnostic matrix construction module is used to integrate the time domain trend, spatial domain distribution and frequency domain characteristics of equipment operation data to construct a three-dimensional diagnostic matrix;

[0067] The collaborative mechanism building module is used to build an edge-cloud collaborative evolution mechanism, deploy a lightweight feature extraction model at the edge, and process high-frequency data in real time; the cloud aggregates knowledge from multiple sites through transfer learning, dynamically updates the global model, and reversely optimizes the edge algorithm.

[0068] In a third aspect, the present application provides an intelligent terminal comprising a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and execute the above-mentioned intelligent management method for energy station operation and maintenance.

[0069] In a fourth aspect, the present application provides a computer-readable storage medium storing a computer program that can be loaded by a processor and execute the above-mentioned intelligent management method for energy station operation and maintenance.

[0070] In summary, compared with the prior art, the above technical solution has the following beneficial effects:

[0071] The intelligent management method and system for energy station operation and maintenance described in this application can maintain high-confidence predictions under small sample and noise data by obtaining the physical operation laws of energy equipment and embedding them into deep learning models, providing a reliable basis for operation and maintenance decisions. A hierarchical feature extraction network is designed for different energy data types, which can effectively identify and enhance key features and achieve the best combination of cross-energy features. A three-dimensional diagnostic matrix is constructed by integrating multi-domain features of equipment operation data, which can comprehensively and accurately diagnose the status of equipment. The edge-cloud collaborative evolution mechanism processes high-frequency data in real time at the edge, aggregates knowledge and optimizes algorithms in the cloud, greatly improving the efficiency and intelligence level of operation and maintenance, reducing operation and maintenance costs, and enhancing the stability and reliability of energy station operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 It is a flow chart of an intelligent management method for energy station operation and maintenance according to an embodiment of the present application. DETAILED DESCRIPTION

[0073] The present application is further described in detail below in conjunction with all the accompanying drawings.

[0074] The present application discloses an intelligent management method and system for energy station operation and maintenance, referring to Figure 1 , an intelligent management method for energy station operation and maintenance, comprising:

[0075] S101. Obtain the physical operating laws of energy equipment.

[0076] Specifically, the management system designs experimental plans for specific energy devices (such as wind turbines, photovoltaic panels, and energy storage batteries). Using sensors and monitoring equipment, it collects real-time operating data (such as temperature, pressure, current, voltage, and speed). Using inductive or deductive reasoning, combined with the laws of physics, it analyzes these experimental data and summarizes the physical operating patterns of the devices under different operating conditions. For example, by analyzing the relationship between a wind turbine's blade speed, wind speed, and generated power, it builds a power output model and records and stores the experimental data in a structured database for subsequent analysis and model training.

[0077] S102. Construct a deep learning model with mechanism constraints.

[0078] Specifically, the management system constructs a mechanism-constrained deep learning model, embedding the physical operating laws of energy equipment into the neural network training process. The deep learning model architecture is selected based on the physical operating laws of the energy equipment. For example, a convolutional neural network (CNN) is used for data with spatial characteristics (such as equipment images and temperature field distributions), while a recurrent neural network (RNN) is used for time series data (such as the changes in equipment operating parameters over time). During model construction, the physical operating laws are embedded into the neural network in the form of constraints. The network depth and width are adjusted based on the complexity of the constraints, and a physical constraint loss term is added to traditional losses (such as mean squared error). For example, the predicted value is forced to satisfy the energy balance equation, and the residual between the predicted value and the theoretical value is calculated as an additional loss. Predictions that violate physical laws are penalized (for example, flow cannot be negative). Weight coefficients are used to balance the contributions of data-driven loss and physical constraint loss. A physical equation solver is introduced at the output layer, for example, by using a custom activation function to constrain the output range (for example, using ReLU to ensure non-negative flow). A physical simulation module (such as finite element method calculation) is connected in parallel with the neural network to jointly optimize parameters.

[0079] S103: Determine the energy data type.

[0080] Specifically, the management system categorizes energy data into different types based on the physical operating patterns of energy equipment and business needs. Energy data types include: structured data, such as equipment operating parameters (temperature, pressure, current, etc.) and equipment status information (switch status, fault codes, etc.); semi-structured data, such as equipment log files and XML-formatted monitoring reports; and unstructured data, such as equipment images, sound signals, and video streams. These different types of data are then preprocessed, including data cleaning (removing noise and outliers), data normalization (unifying data of different dimensions into the same range), and data conversion (converting unstructured data into a structured representation).

[0081] S104. Design a hierarchical feature extraction network to automatically identify and enhance key features.

[0082] Specifically, the management system designs a hierarchical feature extraction network for different energy data types, automatically identifies and strengthens key features, and achieves the optimal combination of cross-energy features through real-time weight adjustment; the hierarchical feature extraction network includes multiple convolutional layers and downsampling layers. The convolutional layers are used to extract local features, and the downsampling layers (such as the maximum pooling layer) are used to reduce the feature dimension and retain important feature information. The management system introduces attention mechanisms at different layers of the network to automatically identify and strengthen key features. For example, the spatial attention mechanism is introduced in the spatial feature extraction layer to focus on the characteristics of the equipment failure area; the temporal attention mechanism is introduced in the temporal feature extraction layer to focus on the trend of equipment operating parameters over time. By designing a dynamic weight adjustment mechanism, the management system adjusts the weights of each layer of the network according to the characteristics of real-time data (such as noise level, data distribution changes, etc.) to achieve the optimal combination of cross-energy features.

[0083] The management system calculates the standard deviation or entropy of real-time data through a sliding window to quantify the noise intensity (for example, high-frequency fluctuation areas have higher noise). It uses statistical indicators (such as KL divergence, difference in moving averages) or online clustering algorithms (such as incremental K-means) to monitor whether the data distribution has shifted. Based on the physical correlation of the multi-dimensional characteristics of energy equipment (such as temperature, pressure, and current), pre-defined feature groups are used to calculate the correlation coefficient within the feature group in real time, and the group boundaries are dynamically adjusted to adapt to data changes. By designing a lightweight sub-network (such as a 1-2 layer fully connected layer), it receives real-time data features as input, outputs weight adjustment coefficients, and assigns independent weight generation units to each network layer. The weight scaling factor of the layer is dynamically generated based on the input features. An attention module is introduced in the skip connection or cross-layer interaction path of the network to calculate the importance scores of the outputs of different layers and dynamically assign fusion weights. For example, if the temperature feature in the current data is abnormal, the weight of the temperature-related feature extraction layer is enhanced.

[0084] S105. Construct a three-dimensional diagnostic matrix.

[0085] Specifically, the management system integrates the time domain trend, spatial domain distribution, and frequency domain characteristics of equipment operation data. Among them, the time domain trend reflects the change pattern of equipment operation parameters over time; the spatial domain distribution reflects the state distribution of the equipment in space (such as temperature field, pressure field, etc.); the frequency domain characteristics reflect the characteristics of equipment operation parameters in the frequency domain (such as vibration frequency, noise frequency, etc.). The management system converts time domain data into frequency domain data through methods such as Fourier transform and wavelet transform; converts discrete spatial data into continuous spatial distribution through methods such as interpolation and gridding, constructs a three-dimensional diagnostic matrix, and uses time domain, spatial domain, and frequency domain characteristics as the three dimensions of the matrix. Use data analysis tools (such as FineBI, MATLAB, etc.) to analyze the three-dimensional matrix, use principal component analysis (PCA) and factor analysis to reduce the dimension and extract key features; use cluster analysis, regression analysis, etc. to establish a diagnostic model to achieve automatic diagnosis and prediction of equipment failures.

[0086] Import the three-dimensional matrix data (e.g., time × sensor type × device instance) collected by device sensors into analysis tools to clarify the physical meaning of the three-dimensional matrix. Combine the time dimension with the device dimension to form a two-dimensional "device-time-feature" table (e.g., each row represents a multi-parameter record for a device at a specific moment). For scenarios with strong temporal dependencies, segment the continuous time data into fixed-length windows (e.g., 10 minutes per window), treating each window as a sample. Z-score normalization (subtracting the mean and dividing by the standard deviation) is performed on each feature column to eliminate dimensionality differences. The principal components, factor scores, and cluster labels after dimensionality reduction are used as model input features, and the training and validation sets are divided into chronological order to avoid future data leakage caused by random splits. The influence of key features on faults is explained using variable importance ranking in random forests or SHAP values (SHapley Additive Explanations). Parameters (e.g., tree depth in random forests and kernel function in SVR) are tuned using grid search or Bayesian optimization.

[0087] S106. Build an edge-cloud collaborative evolution mechanism.

[0088] Specifically, the management system deploys a lightweight feature extraction model at the edge (such as the local server of the energy station, the smart gateway, etc.). The lightweight model can be optimized using model compression technology (such as pruning, quantization, knowledge distillation, etc.) to reduce computational complexity and memory usage. The edge collects equipment operation data in real time, performs feature extraction and preliminary analysis through the lightweight model, and realizes real-time processing of high-frequency data.

[0089] The cloud aggregates knowledge from multiple sites through transfer learning technology. Transfer learning includes feature migration, model migration, parameter migration, and other methods. The appropriate method can be selected according to the data characteristics and model requirements of different sites. The cloud dynamically updates the global model and reversely optimizes the lightweight model on the edge with the updated model parameters. The optimization process can use federated learning technology to protect the data privacy and security of each site.

[0090] Furthermore, the step of obtaining the physical operating law of the energy device specifically includes:

[0091] S101.1. Obtain meteorological satellite data and, in combination with the CAMS atmospheric model, calculate irradiance parameters in real time.

[0092] Specifically, the management system utilizes weather forecast APIs to obtain real-time and forecast weather data. These APIs rely on real-time data collected by weather stations, satellites, radar, and other sources. Weather models analyze and process these data to generate forecasts. In conjunction with the CAMS (Copernicus Atmosphere Monitoring Service) atmospheric model, acquired meteorological satellite data is used to calculate irradiance parameters in real time, including horizontal irradiance (GHI), direct irradiance (DNI), and tilted surface irradiance (GTI). Furthermore, the CAMS atmospheric model provides high-resolution atmospheric composition and meteorological parameters, such as aerosol optical depth, ozone concentration, and cloud cover, which are crucial for irradiance calculations.

[0093] S101.2. Use an improved clear sky model to separate irradiance data under clear and cloudy conditions.

[0094] Specifically, the management system uses improved clear sky models, such as the one based on a gradient boosted decision tree (GBDT), to separate irradiance data for sunny and cloudy conditions. This improved clear sky model considers geographic factors and climatic conditions to determine optimal values for the clear sky model parameters, improving short-term forecast accuracy. The optimal values for the clear sky model parameters are determined sequentially using the least squares method. A gradient boosted decision tree model is then established, using the theoretical daily peak solar irradiance and the air quality index as inputs and the actual daily peak solar irradiance as output. The solar irradiance obtained by the clear sky model is then optimized.

[0095] S101.3. Obtain monitoring information.

[0096] Specifically, the management system obtains monitoring information, which includes terrain elevation data, meteorological historical data, and real-time sensor data; terrain elevation data is obtained through geospatial data cloud, Google Earth, and elevation data points. Terrain elevation data is used to calculate the terrain correction factor of solar irradiance to improve the accuracy of irradiance calculation. Meteorological historical data is obtained from official meteorological bureau websites (such as the official website of the China Meteorological Administration), third-party weather applications, academic research institutions and other channels. Meteorological historical data includes temperature, humidity, precipitation, wind speed, cloud cover, etc., which are used to train prediction models and verify prediction results. Sensors are connected through bus tools (such as I2C, SPI, etc.) to obtain real-time data, and Python scripts are used to collect and store sensor data in real time on edge gateways or in the cloud.

[0097] S101.4. Perform data cleaning and preliminary calibration through the edge gateway.

[0098] Specifically, the management system performs preliminary cleaning of collected data at the edge gateway to remove noise, outliers, and missing values, and smooths time series data using filtering algorithms (such as Kalman filtering and moving average filtering). Sensor data is initially calibrated based on sensor specifications and calibration certificates. Using historical and reference data, the sensor data is calibrated using methods such as linear regression and polynomial fitting.

[0099] S101.5. Build a dynamic virtual weather station.

[0100] Specifically, the management system constructs a dynamic virtual weather station and uses Kalman filtering or Bayesian algorithms to optimize data deviations. It uses the UE (Unreal Engine) virtual engine to create virtual scenes, including terrain, buildings, vegetation, and more. A global 3D map is imported into the UE virtual engine, and 3D models are created for the buildings, terrain, and objects in the virtual scenes. The management system converts weather information obtained from the weather forecast platform into sky, lighting, and particle effect parameter settings within the UE virtual engine, used to simulate weather scenarios such as rain, snow, and fog. Based on real-time weather status information, the management system dynamically adjusts weather parameters in the virtual scene to achieve real-time weather simulation. Kalman filtering or Bayesian algorithms are used to optimize data deviations and improve forecast accuracy.

[0101] S101.6. Analyze feature importance using Random Forest, correct time series deviations using an LSTM network, and output calibrated irradiance values.

[0102] Specifically, the management system uses a random forest algorithm to analyze feature importance and select features with the greatest impact on irradiance prediction. Random forests construct multiple decision trees to calculate feature importance scores and assess their impact on prediction results. The management system employs LSTM (Long Short-Term Memory) networks to correct for time series biases and improve the accuracy of the prediction model. LSTM networks are capable of processing long-term dependencies in time series data and, through memory cells and gating mechanisms, learn complex patterns in time series data.

[0103] S101.7. Combine the gray-level co-occurrence matrix of satellite cloud images with the ARIMA model to predict cloud movement, and use the LSTM network to dynamically correct the numerical weather forecast results.

[0104] Specifically, the management system uses the gray-level co-occurrence matrix of satellite cloud images to analyze cloud texture features and extract features such as cloud shape, size, and texture. The gray-level co-occurrence matrix is a method for describing image texture features. It reflects the image's texture information by calculating the co-occurrence probability between gray levels in the image. The management system uses the ARIMA (Autoregressive Integrated Moving Average) model to predict cloud motion trajectories. The ARIMA model is a time series prediction model composed of three components: autoregression, differentiation, and moving average, and can capture linear dependencies in time series data. The management system uses an LSTM network to dynamically correct numerical weather forecast results to achieve real-time irradiance prediction. The numerical weather forecast results are used as input to the LSTM network, combined with real-time sensor data and satellite cloud image data, to dynamically adjust the prediction results. Using the trained LSTM network, real-time irradiance prediction is achieved, and the prediction results are compared with measured data to evaluate the accuracy and reliability of the prediction model.

[0105] S101.8. Deploy lightweight prediction models on edge servers.

[0106] Specifically, lightweight prediction models are deployed on the edge servers of the management system, and edge gateways integrate data aggregation, anomaly detection, and fault warning functions to reduce the amount of uplink data. Lightweight prediction models are optimized through model compression techniques (such as pruning, quantization, and knowledge distillation) to reduce computing resources and improve response speed. Data aggregation, anomaly detection, and fault warning functions are integrated into edge gateways. The data aggregation function fuses data from multiple sensors to improve data accuracy and reliability. The anomaly detection function uses statistical methods or machine learning algorithms to detect outliers or fault signals in the data. The fault warning function provides early warning of equipment failures or operational risks based on the anomaly detection results.

[0107] Furthermore, after the step of obtaining the physical operating laws of the energy equipment, the method further includes:

[0108] S201: Perform fault logic mining.

[0109] Specifically, the management system conducts fault logic mining based on the physical operating patterns of the energy equipment. This includes fault tree analysis and failure mode and effects analysis (HFMEA). The top event is determined by selecting the fault with the greatest impact on the operation of the energy station as the top event, such as "transformer failure causing a power outage." First, the management system establishes boundary conditions to clarify the system's boundaries, ignoring low-probability events like lightning strikes. Then, through a step-by-step decomposition process, the top event is broken down into intermediate events (such as "transformer winding short circuit" and "transformer insulation damage"), and further into bottom events (such as "winding quality failure" and "insulation material aging"). Logic algebra is used to simplify the fault tree, removing redundant events and logic gates. Finally, the minimum cut set is determined, analyzing all possible combinations that could lead to the top event, calculating the probability of the top event, and evaluating the system's reliability.

[0110] S202. Build a health scoring system.

[0111] Specifically, the management system constructs a health scoring system through scoring methods, which include logistic regression models, random forests, and support vector machines. Among them, the logistic regression model is suitable for analyzing the linear relationship between fault factors and health status, the random forest can handle nonlinear relationships and high-dimensional data, and output feature importance, and the support vector machine performs well in small sample conditions and is suitable for classification and regression tasks. Use historical data to train the model, adjust hyperparameters (such as learning rate, regularization coefficient), use cross-validation to evaluate model performance, and select the optimal model. Input real-time data into the model, calculate the health score (such as 0-100 points), and divide the health level (such as excellent, good, general, poor, and extremely poor) according to the scoring results.

[0112] S203. Construct a multi-dimensional health status model to comprehensively evaluate the health status of the equipment.

[0113] Specifically, the management system collects data from multiple channels such as the SCADA system, sensors, operation and maintenance records, and environmental monitoring equipment. It uses multivariate statistical methods (such as principal component analysis) or machine learning algorithms (such as neural networks) to integrate multi-dimensional data and establish a comprehensive health index (CHI) to reflect the overall health status of the equipment. Based on the data of the equipment's trouble-free operation period, a multi-parameter joint distribution is constructed through the Gaussian mixture model (GMM) or kernel density estimation (KDE), the health status confidence interval is defined, and the baseline mode of key features (such as bearing vibration spectrum characteristics) is stored in the database. Weights are assigned based on the F1-score of the key features on the validation set, and the final health score is comprehensively output. When the equipment is approaching the maintenance cycle, the decision weight of the degradation prediction model is increased.

[0114] S204: Determine the fault factor score and corresponding level.

[0115] Specifically, the management system determines the fault factor score and corresponding level based on the analysis results of the fault logic mining and health scoring system, and realizes a multi-dimensional health assessment. According to the fault logic mining results, the fault factors are divided into categories such as "critical faults", "serious faults", and "general faults". Combined with the results of the health scoring system, a score value is assigned to each fault factor. For example, a critical fault gets 10 points, a serious fault gets 5 points, and a general fault gets 1 point. The highest principle is adopted to determine the level of the fault factor based on its highest score value. For example, a score ≥ 8 points is "high risk", 5-7 points are "medium risk", and 1-4 points are "low risk".

[0116] Furthermore, after the step of designing a hierarchical feature extraction network for different energy data types to automatically identify and enhance key features, the method further includes:

[0117] S301. Dynamically adjust the inverter output strategy.

[0118] Specifically, the management system dynamically adjusts the inverter output strategy based on irradiation parameters and monitoring information, optimizes MPPT tracking efficiency through the PID algorithm, and improves power generation; obtains real-time and historical irradiation parameters through meteorological satellites, ground meteorological stations, etc., optimizes MPPT tracking efficiency through the PID algorithm, and adjusts the operating voltage of the photovoltaic array according to the deviation between the current operating point and the target power point to achieve rapid response; integrates the deviation to eliminate steady-state errors and improve control accuracy; differentiates the rate of change of the deviation to predict future deviations and improve the stability and dynamic performance of the system.

[0119] PID controller parameters are determined based on empirical formulas and on-site debugging. Optimization algorithms (such as genetic algorithms and particle swarm optimization) are then used to automatically tune the PID parameters, improving tuning efficiency and accuracy. Based on irradiation parameters and monitoring information, the maximum power point (MPP) of the photovoltaic array is calculated in real time. The PID controller dynamically adjusts the inverter operating point to ensure it consistently tracks the MPP, improving power generation efficiency. The PID controller's output signal is used to adjust the inverter bridge's operating frequency and duty cycle to control output voltage and current.

[0120] S302: Output device health index (HI).

[0121] Specifically, the management system builds an irradiation-power generation correlation model, inputs historical irradiation data, component attenuation parameters, and cleaning frequency, and outputs the equipment health index (HI), which is used for life prediction and automatically generates operation and maintenance reports and cleaning recommendations. Based on the relationship between historical irradiation data, component attenuation parameters, and power generation, a model is selected for correlation analysis, such as a multiple linear regression model, a random forest regression model, or a gradient boosting tree (GBDT) model. These models can handle the nonlinear relationship between multiple input variables (such as irradiation data, component attenuation parameters, cleaning frequency, etc.) and the output variable (power generation), and have strong fitting and generalization capabilities. The preprocessed data is divided into a training set and a test set, usually in a ratio of 7:3 or 8:2. The training set is used for model training and parameter adjustment, and the test set is used to evaluate the performance and generalization ability of the model.

[0122] Useful features are extracted from the collected raw data as model input variables. In addition to directly using historical irradiation data, component attenuation parameters, and cleaning frequency, derived features such as cumulative irradiation, average irradiation intensity, and cleaning intervals can also be constructed. At the same time, features are screened and optimized to remove those with low correlation with power generation, thereby improving the model's efficiency and accuracy. The selected model is trained using the training set, and the model's loss function on the training set is minimized by adjusting model parameters (such as the learning rate, tree depth, and number of subsamples). During training, cross-validation can be used to avoid overfitting and improve model stability.

[0123] The calculation of the equipment health index (HI) requires comprehensive consideration of multiple factors, including the prediction results of the irradiation-power generation correlation model, component attenuation parameters, cleaning frequency, and the operating status of the equipment (such as temperature, humidity, equipment failure records, etc.). Among them, the difference between the prediction results of the irradiation-power generation correlation model and the actual power generation is an important indicator reflecting the health status of the equipment. The smaller the difference, the better the performance of the equipment and the higher the health index. Based on the factors affecting the health status of the equipment, a complete set of evaluation indicators is established. For example, the following indicators can be set:

[0124] The power generation prediction error rate, i.e. (actual power generation - model predicted power generation) / actual power generation × 100%, reflects the degree of deviation between the model prediction result and the actual power generation.

[0125] The component attenuation rate, i.e. (initial component conversion efficiency - current component conversion efficiency) / initial component conversion efficiency × 100%, reflects the degree of attenuation of component performance.

[0126] The cleaning effect index is calculated based on the change in power generation before and after cleaning, such as (power generation after cleaning - power generation before cleaning) / power generation before cleaning × 100%, reflecting the improvement effect of cleaning on equipment performance.

[0127] Equipment failure frequency, which counts the number of equipment failures within a certain period of time, reflects the reliability of the equipment.

[0128] The weight of each evaluation indicator is determined by using methods such as the analytic hierarchy process (AHP) or the subjective weighting method to reflect the degree of influence of each indicator on the equipment health index. For example, the power generation forecast error rate and the component attenuation rate may have a greater impact on the equipment health index and can be assigned a higher weight; the cleaning effect index and the equipment failure frequency can be assigned a relatively low weight. Based on the established evaluation indicator system and the determined indicator weights, the equipment health index (HI) is calculated using a weighted average method. The equipment health index ranges from 0 to 100. The higher the value, the better the health status of the equipment. The historical health index data, service life, operating environment data (such as temperature, humidity, radiation intensity, etc.) and maintenance records of the equipment are collected as input data for life prediction.

[0129] Cleaning decision rules are established based on factors such as the equipment health index, the degree of component surface contamination (which can be detected through image recognition technology or sensors), historical cleaning frequency data, and local climate conditions (for example, areas with high dust content and low rainfall should have more frequent cleanings). For example, cleaning is recommended when the equipment health index falls below a set threshold (such as 80); cleaning is also recommended when the cleaning interval exceeds a specified period (such as three months).

[0130] Generating cleaning recommendations: Based on established cleaning decision rules, the system automatically analyzes the current status of the equipment and generates corresponding cleaning recommendations. These recommendations include cleaning time, cleaning method (such as manual cleaning, mechanical cleaning, water cleaning, etc.), and cleaning precautions. These recommendations are then sent to the operation and maintenance personnel. The operation and maintenance personnel then arrange cleaning work in a timely manner based on the cleaning recommendations to improve the operating efficiency and lifespan of the equipment.

[0131] Furthermore, the step of automatically generating an operation and maintenance report and cleaning suggestions specifically includes:

[0132] S302.1. Obtain dust accumulation status information.

[0133] Specifically, the management system acquires dust accumulation status information, which includes infrared sensor data, laser scattering data, and image recognition data. It also uses infrared sensors to obtain photovoltaic panel surface temperature data. Infrared sensors can be installed on or near the back of photovoltaic panels to measure panel surface temperature in real time. When dust accumulates on the panel surface, it hinders heat dissipation, causing the surface temperature to rise. Therefore, infrared sensor data can indirectly assess the impact of dust accumulation on panel heat dissipation.

[0134] Using the principle of laser scattering, the intensity of light reflected from the surface of a photovoltaic panel is measured. The laser scattering device transmits a laser beam to the panel surface and measures the intensity of the reflected light. When dust accumulates on the panel surface, the dust scatters the laser beam, reducing the intensity of the reflected light. A camera is used to capture images of the panel surface. The camera can be mounted above or near the panel, capturing real-time images of the panel surface. Using image recognition technology, the image can identify dust accumulation areas and coverage.

[0135] S302.2. Fusion of multimodal data.

[0136] Specifically, the management system fuses multimodal data based on the dust accumulation status information to improve the evaluation accuracy; and preprocesses the infrared sensor data, laser scattering data, and image recognition data to improve the quality and consistency of the data. Features are extracted from the preprocessed data, specifically including extracting temperature anomaly area features from infrared sensor data, extracting reflected light intensity change features from laser scattering data, extracting dust accumulation area shape features from image recognition data, etc. Multimodal data are fused using weighted fusion or machine learning algorithms. Weighted fusion can assign different weights according to the reliability and accuracy of different data sources, and then fuse the weighted data. Machine learning algorithms (such as neural networks) can learn the complex relationships between multimodal data to achieve more accurate data fusion.

[0137] S302.3. Use convolutional neural network to train image recognition model.

[0138] Specifically, the management system uses a convolutional neural network to train an image recognition model, providing real-time output of dust accumulation coverage and power generation efficiency loss rates. The management system collects a large amount of image data from photovoltaic panel surfaces and annotates dust accumulation areas and coverage areas. The image recognition model is trained using a convolutional neural network (CNN), a specialized neural network designed for image recognition that automatically learns features in images and enables efficient image classification and recognition. By adjusting the model structure, hyperparameters, and optimization algorithms, the model's accuracy and generalization capabilities are improved.

[0139] S302.4. Automatically adjust the dust accumulation alarm threshold.

[0140] Specifically, the management system analyzes historical dust accumulation data and meteorological conditions to identify the relationship between the dust accumulation alarm threshold and meteorological conditions. For example, when wind speeds are low and humidity is high, dust is more likely to accumulate on the surface of photovoltaic panels, and in this case, the dust accumulation alarm threshold needs to be lowered. A threshold adjustment model is established to automatically adjust the dust accumulation alarm threshold based on current meteorological conditions (such as wind speed, humidity, and rainfall probability). The threshold adjustment model can use machine learning algorithms such as linear regression and decision trees.

[0141] S302.5. Optimize the cleaning strategy.

[0142] Specifically, the management system optimizes the cleaning strategy with the goal of maximizing net profit. The constraints include:

[0143] Max Profit=∑(P clean ·ΔE·T sun Price)-C clean ·N clean

[0144] Among them, MaxProfit represents the maximum net profit, P clean is the power increase rate after cleaning, ΔE represents the increase in power generation after cleaning, that is, the additional power generation brought by each cleaning, T sun refers to the sunshine time, Price refers to the electricity price, C clean The cost of each cleaning, including labor, materials, equipment loss, etc.

[0145] N clean Refers to the number of cleanings, that is, the total frequency of cleaning the equipment within a period of time.

[0146] With the goal of maximizing net profit, the objective function is defined as the product of the power increase rate after cleaning and the power generation efficiency loss rate, minus the cost of a single cleaning session. Constraints are set, such as the power increase rate after cleaning must not fall below a certain threshold, and the power generation efficiency loss rate must not exceed a certain threshold. These constraints ensure the effectiveness and feasibility of the cleaning strategy. The REINFORCE algorithm is selected as the optimization algorithm. The REINFORCE algorithm is a reinforcement learning algorithm that learns the optimal cleaning strategy by interacting with the environment (such as weather forecasts, electricity price periods, and dust accumulation rates).

[0147] S302.6. Output the optimal cleaning time window.

[0148] Specifically, the management system uses the REINFORCE algorithm, taking as input parameters such as weather forecast (rainfall probability), electricity price period, and dust accumulation rate, to output the optimal cleaning time window. The management system initializes the simulation environment (either randomly or with a preset initial state) and outputs action probabilities and sampled actions (e.g., selecting "delay cleaning by one day") based on the policy network. The action is then executed, the dust accumulation level is updated, and the daily reward is calculated. The system then transitions to the next state, repeating this cycle until the end of the cycle (e.g., 30 days), generating a complete trajectory and reward sequence. A discounted cumulative reward is calculated, and the reward for each step in the trajectory is weighted and summed by a discount factor. The policy network parameters are updated via gradient descent to increase the probability of high-reward actions. A moving average baseline is introduced, calculating the difference between the cumulative reward and the baseline to reduce variance. An entropy term is added to the loss function (to encourage exploration) to prevent premature convergence. After the model makes a decision, a work order is generated and pushed to the operations and maintenance system, including a recommended time, expected reward, and cost details. Operations and maintenance personnel can manually adjust the cleaning time (e.g., delaying the equipment failure period). The system records this feedback for model optimization.

[0149] S302.7. Automatically trigger the cleaning task.

[0150] Specifically, the management system calculates power generation loss costs based on the real-time dust accumulation coverage and power generation efficiency loss rate. This cost is calculated by multiplying the power generation efficiency loss rate by the current electricity price and the installed capacity of the photovoltaic panels. When the real-time calculated power generation loss cost exceeds the single cleaning cost, a cleaning task is automatically triggered and assigned to cleaning equipment or personnel for timely cleaning.

[0151] Furthermore, the step of automatically triggering the cleaning task further includes:

[0152] S302.7.1. Optimize cleaning strategies through reinforcement learning algorithms.

[0153] Specifically, the management system optimizes the cleaning strategy using a reinforcement learning algorithm. It uses the cleanliness-corrected Daily Intake (DNI) as input and incorporates cleaning costs into long-term maintenance costs. The optimal cleaning strategy is determined through simulation and parameter optimization. The management system collects DNI data and cleaning cost data from the PV power plant. DNI data can be obtained through the PV power plant's monitoring system. Cleaning cost data includes manual cleaning costs, equipment wear and tear costs, and other factors. A reinforcement learning model is established, using DNI as input and cleaning costs as output. The model can utilize a deep Q-network (DQN) or a policy gradient algorithm (such as PPO). The optimal cleaning strategy is found through simulation and parameter optimization. The simulation generates different cleaning strategies based on different DNI values and cleaning costs. The strategy is then optimized using a reinforcement learning algorithm to minimize the long-term accumulated cleaning costs.

[0154] S302.7.2. Determine the optimal cleaning time interval using a particle swarm algorithm.

[0155] Specifically, the management system uses a particle swarm algorithm to determine the optimal cleaning interval to maximize net cleaning benefits. Net cleaning benefits are defined as the objective function. Net cleaning benefits equal the increased power generation after cleaning minus the cleaning costs. The particle swarm algorithm optimizes the cleaning intervals to maximize net cleaning benefits. The algorithm randomly initializes a swarm of particles, each representing a cleaning interval. By iteratively updating the particle positions and velocities, the algorithm finds the cleaning interval that maximizes net cleaning benefits.

[0156] S302.7.3. Use K-means clustering algorithm to analyze the weather data of photovoltaic power plants.

[0157] Specifically, the management system uses the K-means clustering algorithm to analyze weather data from the PV power station. Combined with the dust loss coefficient and economic threshold, it dynamically adjusts the cleaning cycle to balance power generation losses with cleaning costs. The K-means clustering algorithm also performs cluster analysis on weather data, grouping similar weather conditions together. The clustering results can be used to predict power generation losses under different weather conditions. The cleaning cycle is dynamically adjusted by combining the dust loss coefficient and economic threshold. The dust loss coefficient reflects the impact of dust on power generation, while the economic threshold determines the cost-benefit balance when cleaning should be performed.

[0158] S302.7.4. Optimize detergent usage and water consumption during the cleaning process using a linear programming model and iterative algorithm.

[0159] Specifically, the management system uses a linear programming model and iterative algorithm to optimize detergent usage and water consumption during the cleaning process, achieving the lowest cost and best cleaning results. Detergent usage and water consumption are defined as the objective function, aiming to minimize resource consumption during the cleaning process. Constraints are set for cleaning performance and cost. The cleaning performance can be evaluated by comparing power generation data before and after cleaning, while the cost includes detergent and water costs. Using the linear programming model and iterative algorithm to find the optimal solution, the linear programming model can find a cleaning solution that minimizes resource consumption based on the objective function and constraints.

[0160] S302.7.5. Use big data analysis and AI technology to compare and evaluate the cleaning effect based on the power generation data before and after cleaning.

[0161] Specifically, the management system uses big data analysis and AI technology to evaluate cleaning effectiveness. By comparing power generation data before and after cleaning, it assesses the impact of cleaning on power generation and continuously optimizes cleaning strategies to improve economic benefits. Based on the cleaning evaluation results, the cleaning strategy can be adjusted, such as the cleaning interval and cleaning method, to increase power generation and reduce cleaning costs.

[0162] S302.7.6. Calculate the cleaning cost-benefit ratio and select the cleaning solution with the highest benefit-to-cost ratio.

[0163] Specifically, the management system calculates the cleaning cost-to-benefit ratio and selects the cleaning solution with the highest benefit-to-cost ratio. Cleaning costs include labor, equipment wear, detergent, and water costs. Cleaning benefits are calculated by multiplying the increased power generation after cleaning by the electricity price. The benefit-to-cost ratio is calculated by dividing the cleaning benefits by the cleaning costs, and the cleaning solution with the highest benefit-to-cost ratio is selected. The benefit-to-cost ratio is calculated for different cleaning solutions, and the solution with the highest ratio is selected as the optimal cleaning solution.

[0164] Furthermore, after the step of automatically triggering the cleaning task, the method further includes:

[0165] S401 uses a tightly coupled solution of lidar, visual camera and inertial measurement unit to build a three-dimensional point cloud map through front-end laser mileage and back-end map optimization technology.

[0166] Specifically, the management system uses a tightly coupled solution of lidar, visual cameras, and inertial measurement units to construct a three-dimensional point cloud map through front-end laser odometry and back-end graph optimization technology. The lidar, visual cameras, and inertial measurement unit (IMU) are integrated on the mobile robot to ensure temporal and spatial synchronization between the sensors. The front-end laser odometry is implemented using lidar data, and the robot's posture changes are estimated by matching point cloud data at different times. Back-end graph optimization technology, combined with a closed-loop detection algorithm, globally optimizes the pose estimation of the front-end laser odometry to improve pose estimation accuracy and map consistency. The lidar and IMU data are integrated to construct a three-dimensional point cloud map. The IMU provides high-frequency posture change information to compensate for the lidar's posture estimation when the scanning frequency is low or when it fails for a short time.

[0167] S402. Collect point cloud data through laser SLAM equipment and generate a three-dimensional model with color information by combining visual texture mapping technology.

[0168] Specifically, the management system uses laser SLAM equipment to collect point cloud data, acquiring 3D point cloud data from the environment. Combined with visual texture mapping technology, this visual information is integrated into the 3D model, generating a color-enhanced 3D model to enhance the model's visual quality and recognition capabilities. Point cloud data and visual images at the same moment are matched based on the timestamps recorded by the device. If there are errors in time synchronization, keyframes are manually or automatically aligned using a feature matching algorithm. The camera coordinate system is aligned with the laser SLAM coordinate system, and coordinate transformations are performed using device calibration parameters or common feature points (such as corners or landmarks) to ensure that the spatial positions of the point cloud and image are consistent. For each laser point cloud, the pixel color value of the corresponding image is assigned to that point based on its spatial coordinates and the camera's viewpoint. For example, a projection transformation is used to project the point cloud onto the corresponding image, find the RGB value of the nearest neighbor pixel, and add color to the point cloud. This process is repeated for each scanned point cloud and image frame to ensure that the visual texture is integrated throughout the entire point cloud dataset. Modeling software (such as Meshlab or Blender) is used to convert the colored point cloud into a triangular mesh model to generate a continuous surface structure. Repair holes in the model (e.g., missing point clouds due to occlusion), manually adjust texture misalignment, and add lighting effects to simulate a real-world environment to enhance the model's visual layering. Save the final model in a common format (e.g., OBJ, PLY, STL) for easy access in management systems for visualization or analysis.

[0169] S403: For dynamic environments, a closed-loop detection algorithm and a dynamic window method are introduced to update the map in real time.

[0170] Specifically, the management system incorporates a closed-loop detection algorithm and a dynamic window method to update maps in real time for dynamic environments. The closed-loop detection algorithm identifies and corrects accumulated drift errors, improving map accuracy and consistency. The dynamic window method handles dynamic obstacles and the robot's dynamic constraints, enabling real-time obstacle avoidance and path planning.

[0171] S404. Classify the clean area according to the complexity of the site terrain.

[0172] Specifically, the management system divides the site into flat, complex, and dynamic areas based on the complexity of the site's terrain. The flat area has simple terrain and is suitable for routine cleaning operations; the complex area has complex terrain with obstacles and slope changes; and the dynamic area contains moving objects, requiring real-time perception and obstacle avoidance.

[0173] S405: For the flat area, a rasterized map is used to mark the regular cleaning path; for the complex area, an octree map is constructed to mark the obstacle collision risk through three-dimensional voxels; for the dynamic area, SLAM is used to perceive moving objects in real time to trigger dynamic map updates.

[0174] Specifically, for flat areas, the management system uses a rasterized map to divide the environment into equally sized grid cells and mark regular cleaning paths, allowing the robot to clean along the intended path. For complex areas, an octree map is constructed, dividing the space into cubic cells. Obstacle collision risks are marked using 3D voxels, providing the robot with obstacle avoidance information. For dynamic areas, SLAM (Simultaneous Localization and Mapping) is used to detect moving objects, such as people and vehicles, in real time, triggering dynamic map updates to maintain real-time consistency between the map and the environment.

[0175] S406. Integrate Euclidean distance, Manhattan distance and height influence factor to define energy consumption cost function.

[0176] Specifically, the management system integrates Euclidean distance, Manhattan distance and height impact factor to define the energy consumption cost function:

[0177] f(n)=g(n)+λ1·h distance (n)+λ2·h height (n)

[0178] Where g(n) represents the actual energy consumption accumulated from the starting point to the current node n, which can be calculated through device sensors (such as motor power consumption monitoring) or preset energy consumption models (such as energy consumption per unit distance moved); h distance (n) represents the straight-line distance from the current node n to the end point; h height (n) represents the slope at node n. The larger the slope, the higher the h height The higher the (n) value, the more it reflects the impact of climbing or descending on energy consumption; λ1 and λ2 are weight coefficients. When the impact of terrain on energy consumption is greater than that of distance, λ2 is set to a larger value and optimized through debugging (for example, adjusting the weight in the test area to make the path planning result meet the expected energy consumption).

[0179] S407: Adaptively adjust the search step size according to the regional complexity.

[0180] Specifically, the management system adaptively adjusts the search step size based on the complexity of the area. Flat areas use large step sizes to accelerate the search and improve planning efficiency; complex areas use small step sizes for precise obstacle avoidance and ensure path safety. The slope variance of the current area is calculated using the height map. A small variance indicates a flat area; a large variance or the presence of multiple height mutation points indicates a complex area. The number of obstacles per unit area is counted (e.g., through lidar or visual recognition). A small number indicates a flat area, while a large number indicates a complex area. With each planning step, the complexity of the next area is reassessed and the step size is dynamically adjusted.

[0181] S408: Smoothing the broken line path using a quasi-uniform B-spline curve.

[0182] Specifically, the management system uses quasi-uniform B-spline curves to smooth broken-line paths, reducing the number of turns and energy loss, and improving the feasibility and efficiency of the path. The initial broken-line path is generated using traditional algorithms (such as A* and Dijkstra), resulting in a series of discrete nodes (a sequence of path point coordinates). Key nodes in the broken-line path are selected according to certain rules (such as every few nodes) as control points for the B-spline curve to ensure that the curve reflects the path trend. Using the quasi-uniform B-spline algorithm, the curve equation is calculated based on the control points and the curve order, generating a smooth curve that passes through or approaches the control points, replacing the original broken-line path. Finally, the smoothed path is checked for conflicts with obstacles (e.g., collision detection is performed on points on the sampling curve). If so, the control points are adjusted or the original broken-line nodes are re-smoothed until the path is feasible.

[0183] S409. Combining spiral scanning with a reciprocating path, achieving non-overlapping coverage through regional decomposition.

[0184] Specifically, the management system selects the center point or an unobstructed point near the entrance in each sub-area as the starting point of the spiral. Starting from the starting point, the spiral line is expanded outward at fixed increments (such as 0.8 times the cleaning width of the equipment). The area covered by each spiral must ensure that adjacent paths have partial overlap (such as a 10% overlap rate) to avoid missed scans. When the spiral scan approaches the boundary of the sub-area (1-2 meters away from the boundary through map coordinates or real-time monitoring by sensors), the reciprocating path is triggered. Move in a straight line along the edge of the sub-area, and turn back after reaching one end. The width of each movement is consistent with the cleaning width of the equipment. For example, if the sub-area is rectangular, when the spiral scan reaches the short side boundary, it switches to reciprocating movement along the long side until the entire sub-area is covered.

[0185] S410 , dynamically adjusting the cleaning order based on a Markov decision process (MDP).

[0186] Specifically, the management system establishes a cleaning order adjustment model based on the Markov decision process (MDP), decomposing the cleaning task into multiple states, each of which corresponds to a cleaning area. The state is defined as a cleaning area, and the actions include movement and cleaning. The movement action moves the robot from the current area to the target area; the cleaning action cleans the target area. The reward function is designed based on the cleaning effect and energy consumption. The cleaning effect is evaluated by comparing the power generation data before and after cleaning; energy consumption includes movement energy consumption and cleaning energy consumption. The reward function encourages the robot to prioritize high-pollution areas and reduce the overall movement distance and energy consumption. The cleaning order is optimized through the MDP algorithm to reduce the overall movement distance and energy consumption. The MDP algorithm can dynamically adjust the cleaning order according to the current state and environmental information to achieve the optimal cleaning strategy.

[0187] Based on the above method, the embodiment of the present application also discloses an intelligent management system for energy station operation and maintenance. An intelligent management system for energy station operation and maintenance includes:

[0188] Physical operation law acquisition module, used to obtain the physical operation law of energy equipment;

[0189] The learning model construction module is used to build a deep learning model with mechanism constraints, embedding the physical operation laws of energy equipment into the neural network training process, so that the model can maintain high-confidence predictions in scenarios with small samples and noisy data;

[0190] An energy type determination module, configured to determine the type of energy data according to the physical operating rules of the energy equipment;

[0191] Key feature extraction module, which is used to design a hierarchical feature extraction network for different energy data types, automatically identify and enhance key features, and achieve the optimal combination of cross-energy features through real-time weight adjustment;

[0192] The diagnostic matrix construction module is used to integrate the time domain trend, spatial domain distribution and frequency domain characteristics of equipment operation data to construct a three-dimensional diagnostic matrix;

[0193] The collaborative mechanism building module is used to build an edge-cloud collaborative evolution mechanism, deploy a lightweight feature extraction model at the edge, and process high-frequency data in real time; the cloud aggregates knowledge from multiple sites through transfer learning, dynamically updates the global model, and reversely optimizes the edge algorithm.

[0194] An embodiment of the present application also discloses an intelligent terminal, which includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed by the intelligent management method for energy station operation and maintenance as described above.

[0195] The present application also discloses a computer-readable storage medium. The computer-readable storage medium stores a computer program capable of being loaded by a processor and executing the intelligent management method for energy station operation and maintenance described above. The computer-readable storage medium includes, for example, a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, among other media capable of storing program code.

[0196] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the scope of protection of the invention. Obviously, the embodiments described are only some embodiments of the present invention, rather than all embodiments. Based on these embodiments, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in this field can still combine, add, delete or make other adjustments to the features in the various embodiments of the present invention according to the circumstances without conflict, without making creative work, so as to obtain different other technical solutions that do not deviate from the concept of the present invention in essence, and these technical solutions also fall within the scope of protection of the present invention.

Claims

1. An intelligent management method for energy station operation and maintenance, characterized in that: The following steps are involved: Obtain the physical operating laws of energy equipment; Build a mechanism-constrained deep learning model that embeds the physical operating laws of energy equipment into the neural network training process, enabling the model to maintain high-confidence predictions in scenarios with small samples and noisy data; Determining the energy data type according to the physical operating laws of the energy equipment; Design a hierarchical feature extraction network for different energy data types to automatically identify and enhance key features, and achieve the optimal combination of cross-energy features through real-time weight adjustment; Integrate the time domain trend, spatial distribution and frequency domain characteristics of equipment operation data to build a three-dimensional diagnostic matrix; Build an edge-cloud co-evolution mechanism, deploy lightweight feature extraction models at the edge, and process high-frequency data in real time; The cloud aggregates knowledge from multiple sites through transfer learning, dynamically updates the global model, and reversely optimizes the edge algorithm.

2. The intelligent management method for energy station operation and maintenance according to claim 1 is characterized in that: After obtaining the physical operating laws of energy equipment, it also includes: Perform fault logic mining based on the physical operation rules of the energy equipment, wherein the fault logic mining includes fault tree analysis and failure mode and effects analysis (HFMEA); Constructing a health scoring system through scoring methods, including logistic regression model, random forest, and support vector machine; Build a multi-dimensional health status model to comprehensively evaluate the health status of equipment; According to the analysis results of the fault logic mining and health scoring system, the fault factor score value and the corresponding level are determined to achieve a multi-dimensional health assessment.

3. The intelligent management method for energy station operation and maintenance according to claim 1 is characterized in that: The step of obtaining the physical operating laws of the energy device specifically includes: Acquire meteorological satellite data and combine it with the CAMS atmospheric model to calculate irradiance parameters in real time, including horizontal irradiance (GHI), direct irradiance (DNI), and tilted irradiance (GTI); An improved clear sky model is used to separate irradiance data under clear and cloudy conditions to improve short-term forecast accuracy; Acquiring monitoring information, including terrain elevation data, historical meteorological data, and real-time sensor data; Data cleaning and preliminary calibration are performed through the edge gateway to reduce redundant data transmission; Build a dynamic virtual weather station and use Kalman filtering or Bayesian algorithm to optimize data deviation; The feature importance is analyzed through random forest, the time series deviation is corrected through LSTM network, and the calibrated irradiation value is output; Combining the gray-level co-occurrence matrix of satellite cloud images with the ARIMA model to predict cloud movement, the LSTM network is used to dynamically correct the numerical weather forecast results to achieve irradiance prediction; Edge servers deploy lightweight prediction models, and edge gateways integrate data aggregation, anomaly detection, and fault warning functions to reduce the amount of uplink data.

4. The intelligent management method for energy station operation and maintenance according to claim 3 is characterized in that: After designing a hierarchical feature extraction network for different energy data types to automatically identify and enhance key features, the following steps are also included: Based on irradiation parameters and monitoring information, the inverter output strategy is dynamically adjusted, and the MPPT tracking efficiency is optimized through the PID algorithm to increase power generation; Build an irradiation-energy production correlation model, input historical irradiation data, component attenuation parameters, and cleaning frequency, and output the equipment health index (HI) for life prediction and automatic generation of operation and maintenance reports and cleaning recommendations.

5. The intelligent management method for energy station operation and maintenance according to claim 4 is characterized in that: The steps of automatically generating an operation and maintenance report and cleaning suggestions specifically include: Acquiring dust accumulation status information, wherein the dust accumulation status information includes infrared sensor data, laser scattering data, and image recognition data; According to the dust accumulation status information, multimodal data is integrated to improve the assessment accuracy; Use convolutional neural networks to train image recognition models and output dust accumulation coverage and power generation efficiency loss rate in real time; Automatically adjust dust accumulation alarm threshold based on historical data and meteorological conditions; Optimize the cleaning strategy with the goal of maximizing net profit. The constraints include: Max Profit=∑(P clean ·ΔE·T sun Price)-C clean ·N clean Among them, MaxProfit represents the maximum net profit, P clean is the power improvement rate after cleaning, ΔE represents the increase in power generation after cleaning, T sun refers to the sunshine time, Price refers to the electricity price, C clean is the cost of each cleaning, N clean Refers to the number of cleanings; Using the REINFORCE algorithm, input parameters such as weather forecast (rainfall probability), electricity price period, dust accumulation rate, etc., and output the optimal cleaning time window; When the real-time calculated power generation loss cost is greater than or equal to the single cleaning cost, the cleaning task is automatically triggered.

6. The intelligent management method for energy station operation and maintenance according to claim 1 is characterized in that: The step of automatically triggering the cleaning task also includes: The cleaning strategy is optimized through a reinforcement learning algorithm, taking the cleanliness-corrected daily intake (DNI) as input and incorporating cleaning costs into long-term maintenance costs. The optimal cleaning strategy is achieved through simulation and parameter optimization. The particle swarm algorithm is used to determine the optimal cleaning time interval, thereby improving the net cleaning benefit; Using the K-means clustering algorithm to analyze weather data from photovoltaic power plants, combined with the dust loss coefficient and economic threshold, the cleaning cycle is dynamically adjusted to balance power generation loss and cleaning costs. Through linear programming models and iterative algorithms, the amount of detergent used and water consumption during the cleaning process are optimized to achieve the lowest cost and the best cleaning effect; Utilizing big data analysis and AI technology, we evaluate cleaning effectiveness based on power generation data before and after cleaning, and continuously optimize cleaning strategies to improve economic benefits. By calculating the cleaning cost and benefit ratio, select the cleaning plan with the highest benefit to cost ratio.

7. The intelligent management method for energy station operation and maintenance according to claim 1 is characterized in that: After the step of automatically triggering the cleaning task, the method further includes: Adopting a tightly coupled solution of LiDAR, visual cameras, and inertial measurement units, it constructs a 3D point cloud map through front-end laser odometer and back-end map optimization technology. Use laser SLAM equipment to collect point cloud data and combine it with visual texture mapping technology to generate a three-dimensional model with color information; For dynamic environments, closed-loop detection algorithms and dynamic window methods are introduced to update maps in real time; Classify the clean area into levels according to the complexity of the site terrain, including flat area, complex area, and dynamic area; For the flat area, a rasterized map is used to mark the conventional cleaning path; for the complex area, an octree map is constructed to mark the obstacle collision risk through three-dimensional voxels; for the dynamic area, SLAM is used to perceive moving objects in real time to trigger dynamic map updates; Combining Euclidean distance, Manhattan distance and height influence factor, we define the energy cost function: f(n)=g(n)+λ1·h distance (n)+λ2·h height (n) Among them, g(n) represents the actual energy consumption accumulated from the starting point to the current node n, h distance (n) represents the straight-line distance from the current node n to the end point, h height (n) represents the slope at node n, λ1 and λ2 are weight coefficients; Adaptively adjust the search step size based on the complexity of the area. Use a large step size to accelerate the search in flat areas, and a small step size to avoid obstacles in complex areas. Quasi-uniform B-spline curves are used to smooth the broken line path to reduce the number of turns and energy loss; Combining spiral scanning with reciprocating paths, it achieves non-repetitive coverage through regional decomposition and reduces ineffective movement; The cleaning order is dynamically adjusted based on the Markov decision process (MDP), giving priority to high-pollution areas and reducing the overall moving distance.

8. An intelligent management system for energy station operation and maintenance, characterized by: include: Physical operation law acquisition module, used to obtain the physical operation law of energy equipment; The learning model construction module is used to build a deep learning model with mechanism constraints, embedding the physical operation laws of energy equipment into the neural network training process, so that the model can maintain high-confidence predictions in scenarios with small samples and noisy data; An energy type determination module, configured to determine the type of energy data according to the physical operating rules of the energy equipment; Key feature extraction module, which is used to design a hierarchical feature extraction network for different energy data types, automatically identify and enhance key features, and achieve the optimal combination of cross-energy features through real-time weight adjustment; The diagnostic matrix construction module is used to integrate the time domain trend, spatial domain distribution and frequency domain characteristics of equipment operation data to construct a three-dimensional diagnostic matrix; A collaborative mechanism building module is used to build an edge-cloud collaborative evolution mechanism, deploy a lightweight feature extraction model at the edge, and process high-frequency data in real time; The cloud aggregates knowledge from multiple sites through transfer learning, dynamically updates the global model, and reversely optimizes the edge algorithm.

Images