An intelligent operation and maintenance management system based on new energy equipment

Abstract

The application discloses a kind of based on new energy equipment intelligent operation and maintenance management system, belong to new energy equipment operation and maintenance management technical field.The system includes data acquisition unit, state evaluation unit, risk prediction unit, maintenance decision unit and interactive display unit.Data acquisition unit is fused according to equipment identification and time stamp to multi-source heterogeneous operation and maintenance data, generates standardized operation and maintenance dataset;State evaluation unit calls pre-trained graph neural network to model the spatio-temporal coupling relationship between devices, obtains device health state atlas;Risk prediction unit uses time series convolution network to deduce the future operation trend of the device based on the above atlas, generates fault risk prediction results with uncertainty interval;Maintenance decision unit carries out multi-objective optimization between operation and maintenance cost and power generation income through reinforcement learning agent, outputs equipment maintenance strategy suggestion;Interactive display unit converts maintenance strategy suggestion into three-dimensional dynamic visualization interface, responds to operation and maintenance personnel operation instruction and carries out linkage adjustment.

Description

Technical Field

[0001] This invention relates to the field of operation and maintenance management technology for new energy equipment, specifically to an intelligent operation and maintenance management system for new energy equipment. Background Technology

[0002] Distributed renewable energy power stations include various equipment such as wind turbines, photovoltaic modules, and energy storage batteries. Their operation and maintenance management requires handling heterogeneous data from multiple sources, including vibration, current, voltage, temperature, and meteorological data. Existing operation and maintenance management technologies typically employ isolated data processing pipelines, cleaning and storing data from different sources separately, and then extrapolating equipment health based on preset thresholds or simple trends. This approach struggles to characterize the interactions between devices through electrical connections and spatial locations when dealing with large-scale equipment clusters. Condition assessments are limited to the historical operating curves of individual devices, resulting in an inability to provide accurate health evaluations in the early stages of cascading failures or when performance degradation is latent.

[0003] In terms of fault prediction, conventional methods often employ deterministic prediction models, outputting only a single numerical value for remaining lifespan or failure probability, lacking a quantitative expression of the uncertainty of the prediction results. When renewable energy equipment experiences drastic changes in operating conditions due to weather fluctuations or grid dispatch, this deterministic output cannot provide maintenance personnel with a level of confidence in risk assessment, easily leading to false alarms or missed alarms. In the maintenance decision-making stage, plans are usually generated based on fixed maintenance cycles or simple cost constraints. This fails to simultaneously address the conflicting objectives of maximizing power generation revenue and minimizing maintenance costs when maintenance resource inventory changes dynamically. The resulting maintenance plans are economically inefficient and ill-suited to the reality of frequent fluctuations in the marginal revenue of renewable energy power generation.

[0004] Furthermore, the interactive display layer often uses two-dimensional reports or static topology diagrams, making it difficult for maintenance personnel to intuitively perceive the spatial distribution and internal details of equipment health status, resulting in high operational feedback delays. In summary, existing technologies need to address how to obtain fault risk prediction results with quantified uncertainties based on fully extracting the spatiotemporal coupling characteristics between distributed renewable energy devices, and how to optimize multi-objective maintenance strategies to achieve a dynamic balance between maintenance costs and power generation revenue. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent operation and maintenance management system for new energy equipment, in order to solve the problems in the existing technology where equipment condition assessment ignores the spatiotemporal coupling relationship, resulting in a lack of uncertainty quantification in fault prediction, and where maintenance decisions cannot take into account the multi-objective optimization of operation and maintenance costs and power generation revenue.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides an intelligent operation and maintenance management system for new energy equipment, including a data acquisition unit, a status assessment unit, a risk prediction unit, a maintenance decision-making unit, and an interactive display unit. The data acquisition unit is used to collect multi-source heterogeneous operation and maintenance data from distributed new energy equipment, and to associate and fuse this multi-source heterogeneous operation and maintenance data according to equipment identifiers and timestamps to generate a standardized operation and maintenance dataset, thereby eliminating the problems of data source differences and spatiotemporal asynchrony, and providing a unified and standardized data foundation for subsequent analysis. The status assessment unit is used, based on the standardized operation and maintenance dataset, to call a pre-trained graph neural network to model the spatiotemporal coupling relationship between new energy equipment, extract the spatiotemporal dependency features between equipment and map them into health score vectors, generating an equipment health status map, thereby accurately capturing the overall health status of the equipment cluster and its mutual influences. The risk prediction unit is used, based on the equipment health status map, to use a temporal convolutional network to extrapolate the future operating trend of the equipment, generating a fault risk prediction result with an uncertainty interval, which can quantify the credibility of the prediction result and provide richer information for early warning decisions. The maintenance decision-making unit receives the fault risk prediction results and performs multi-objective optimization between operation and maintenance costs and power generation revenue through a reinforcement learning agent, outputting equipment maintenance strategy suggestions to effectively control operation and maintenance costs while ensuring power generation revenue. The interactive display unit constructs a three-dimensional dynamic visualization interface based on the equipment maintenance strategy suggestions and responds to operation and maintenance personnel's commands, adjusting the equipment parameters in the three-dimensional dynamic visualization interface accordingly. This allows operation and maintenance personnel to intuitively grasp the equipment status and maintenance plan, improving interaction efficiency and decision-making transparency.

[0007] In a preferred embodiment of the present invention, the specific process by which the data acquisition unit correlates and fuses the multi-source heterogeneous operation and maintenance data to generate a standardized operation and maintenance dataset is as follows: Vibration data, current and voltage data, temperature data, and meteorological environmental data of wind turbines, photovoltaic modules, and energy storage batteries are acquired; the vibration data, current and voltage data, temperature data, and meteorological environmental data are grouped according to the same equipment identifier, and alignment and interpolation processing is performed within each group according to the same timestamp; the multi-source data after alignment and interpolation processing is converted into a standardized field sequence in a unified format, which serves as the standardized operation and maintenance dataset. This method effectively solves the inconsistency problem of multi-source data in terms of acquisition frequency, data format, and spatiotemporal reference.

[0008] Preferably, the specific process by which the state assessment unit calls a pre-trained graph neural network to model the spatiotemporal coupling relationship between new energy devices and obtain the device health status map is as follows: The operating parameter sequence of each device in the standardized operation and maintenance dataset is obtained, and an initial graph structure is constructed with individual devices as nodes and electrical connections between devices as edges; the operating parameter sequence is input as node features into the graph neural network, and the spatiotemporal dependency features of each device at adjacent time steps are extracted through multi-layer convolution operations of the graph neural network; the spatiotemporal dependency features are mapped to health score vectors, and the health score vectors are bound to the initial graph structure to generate the device health status map. Modeling the complex spatiotemporal coupling relationship between devices using a graph neural network can significantly improve the accuracy and robustness of state assessment.

[0009] Preferably, the specific process by which the risk prediction unit uses a temporal convolutional network to extrapolate the future operating trend of the equipment based on the equipment health status map and generate a fault risk prediction result with an uncertainty interval is as follows: Extract the health score sequence of each device and the correlation strength sequence between adjacent devices from the equipment health status map; concatenate the health score sequence and the correlation strength sequence into a multi-channel temporal input matrix, and input the multi-channel temporal input matrix into the temporal convolutional network; calculate the health score prediction interval for multiple future prediction times through the dilated convolutional layer and residual connection layer in the temporal convolutional network, and mark the times when the health score prediction interval exceeds a preset threshold as risk times, thereby generating the fault risk prediction result. This method utilizes dilated convolution to expand the receptive field to capture long-term dependencies, and residual connections ensure the stability of deep network training, making the risk extrapolation results more reliable.

[0010] Preferably, the specific process by which the maintenance decision-making unit, through a reinforcement learning agent, performs multi-objective optimization between operation and maintenance costs and power generation revenue to output equipment maintenance strategy recommendations is as follows: The risk time and corresponding risk equipment identifier marked in the fault risk prediction results are obtained; the risk equipment identifier, current operation and maintenance resource inventory, and future power generation prediction curves are input into the reinforcement learning agent as the state space; the reinforcement learning agent selects maintenance actions in the action space, the maintenance actions including maintenance time, maintenance type, and maintenance resource allocation, and outputs the optimal equipment maintenance strategy recommendation by maximizing the cumulative reward function, whereby the cumulative reward function is negatively correlated with operation and maintenance costs and positively correlated with power generation revenue. By constructing a reward function directly related to costs and revenues, maintenance decisions can dynamically balance multiple objectives, achieving optimal resource allocation.

[0011] Preferably, the interactive display unit constructs a three-dimensional dynamic visualization interface based on the equipment maintenance strategy recommendations and adjusts in response to the operation instructions of maintenance personnel as follows: Each device in the equipment maintenance strategy recommendations is used to generate a corresponding three-dimensional model in a three-dimensional geographic coordinate system, and the health score vector from the equipment health status map is overlaid on the three-dimensional model; clicks or selections on the three-dimensional model by maintenance personnel are received, and corresponding detailed operating parameters are retrieved from the standardized maintenance dataset and displayed in a floating window; the display status of the three-dimensional model is updated synchronously according to the parameter adjustment instructions input by the maintenance personnel, and the status assessment unit is triggered to recalculate the health score of the associated equipment. Furthermore, the interactive display unit also supports maintenance personnel to perform sectioning operations on the three-dimensional model and displays real-time temperature field and stress distribution cloud maps of key internal components at the sectioning surface, enabling maintenance personnel to gain a deeper understanding of the internal operating status of the equipment and promptly detect potential anomalies.

[0012] As a preferred embodiment of the present invention, the system further includes an adaptive calibration unit, used to receive maintenance operation records actually performed by maintenance personnel from the interactive display unit, compare the maintenance operation records with the equipment maintenance strategy suggestions, and calculate a deviation vector; use the deviation vector to perform online gradient updates on the node weights of the graph neural network and the reward function parameters of the reinforcement learning agent, generating a calibrated graph neural network and a calibrated reinforcement learning agent; and send the calibrated graph neural network and the calibrated reinforcement learning agent to the state assessment unit and the maintenance decision unit respectively, replacing the original model parameters. This closed-loop feedback calibration mechanism enables the system to continuously learn from real operation and maintenance behaviors, constantly improving the accuracy and applicability of state assessment and maintenance decisions.

[0013] More preferably, the online gradient update employs an elastic weight consolidation regularization term, which limits the variation of the node weights and the reward function parameters during the update process by applying deviation penalties to the model parameters, ensuring that the system retains the original learned knowledge while adapting to new data and new scenarios, and maintaining the long-term stability of model performance.

[0014] In specific implementation, the adaptive calibration unit uses the deviation vector for online gradient updates as follows: It extracts the actual maintenance time and actual maintenance resource consumption from the maintenance operation record, and extracts the suggested maintenance time and suggested maintenance resource consumption from the equipment maintenance strategy suggestions; it combines the time difference between the actual maintenance time and the suggested maintenance time, and the resource difference between the actual maintenance resource consumption and the suggested maintenance resource consumption, into the deviation vector; it uses the deviation vector as a supervision signal to calculate the first loss gradient of the graph neural network output layer and the second loss gradient of the reinforcement learning agent Q-network; and it updates the node weights and the reward function parameters based on the first and second loss gradients, respectively. This update method directly uses the actual operation and maintenance deviation as the learning target, enabling the model to converge more quickly to the optimal strategy that conforms to the actual situation on site.

[0015] As another preferred embodiment of the present invention, the system further includes a data integrity verification unit, used to detect whether there are missing data intervals in the timestamp sequence corresponding to each device identifier in the standardized operation and maintenance dataset after the data acquisition unit generates the dataset; for device identifiers with missing data intervals, to find data of other device identifiers spatially adjacent to the device identifier at the same timestamp, and to fill the missing data intervals using spatial interpolation methods to generate a completed standardized operation and maintenance dataset; and to send the completed standardized operation and maintenance dataset to the status assessment unit. This unit effectively improves the completeness of the input data and avoids inaccurate subsequent analysis due to missing data.

[0016] As another preferred embodiment of the present invention, the system further includes a confidence assessment unit, used to obtain the uncertainty interval in the fault risk prediction result generated by the risk prediction unit, and compare the uncertainty interval with a preset confidence threshold; when the uncertainty interval exceeds the confidence threshold, the data acquisition unit is triggered to increase the data sampling frequency of the corresponding device, and the new data sampled at the high frequency is re-inputted into the risk prediction unit for secondary deduction to obtain an updated fault risk prediction result; the portion of the updated fault risk prediction result with an uncertainty interval lower than the confidence threshold is marked as a high-confidence risk event and directly pushed to the interactive display unit for highlighting. By dynamically increasing the data sampling frequency and iterative prediction, the uncertainty interval of the prediction is effectively compressed, making the early warning of high-risk events more clear and reliable, and facilitating maintenance personnel to take timely and targeted measures.

[0017] This invention, through the collaborative operation of its various units, organically combines multi-source data fusion, spatiotemporal state assessment, uncertainty risk simulation, multi-objective maintenance decision-making, and three-dimensional visualization interaction to form an intelligent closed loop from data acquisition to maintenance execution, significantly improving the automation level and overall benefits of new energy equipment operation and maintenance management.

[0018] The technical effects and advantages provided by the present invention in the above technical solution are as follows: A graph neural network is employed to model the spatiotemporal coupling relationships among new energy equipment. A graph structure is constructed using individual equipment as nodes and electrical connections between equipment as edges. The operating parameter sequences of each equipment are used as node feature inputs. Multi-layer convolutional operations extract the spatiotemporal dependency features of each equipment at adjacent time steps and map them into health score vectors, generating an equipment health status map. From the map, health score sequences and correlation strength sequences between adjacent equipment are extracted and concatenated into a multi-channel temporal input matrix. This matrix is ​​input into a temporal convolutional network, where dilated convolutional layers and residual connection layers are used to calculate the health score prediction intervals for multiple future prediction times. Moments exceeding a preset threshold are marked as risk moments, resulting in a fault risk prediction result with an uncertainty interval. This approach captures the spatial coupling effects and temporal evolution patterns between equipment, presenting fault risks as probability intervals. It overcomes the problem that deterministic predictions of a single equipment cannot measure confidence levels, allowing maintenance personnel to make differentiated responses based on the breadth of risk uncertainty, reducing the probability of misjudgments and omissions.

[0019] By employing a reinforcement learning agent to perform multi-objective optimization between operation and maintenance costs and power generation revenue, the system outputs equipment maintenance strategy suggestions. It obtains risk moments and risk equipment identifiers from fault risk prediction results. These risk equipment identifiers, current operation and maintenance resource inventory, and future power generation forecast curves are input into the reinforcement learning agent's state space. The agent selects maintenance actions from an action space that includes maintenance time, maintenance type, and maintenance resource allocation, maximizing the cumulative reward function. This function is negatively correlated with operation and maintenance costs and positively correlated with power generation revenue. This approach automatically seeks a dynamic balance between reducing maintenance resource consumption and increasing power generation revenue, generating maintenance timing and methods that adapt to changes in resource inventory and electricity price fluctuations. It avoids the shortcomings of fixed rules or single-objective optimization under complex constraints, which cannot simultaneously consider economic efficiency, thus improving the efficiency of operation and maintenance resource utilization and overall power generation revenue. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0021] Figure 1 This is a structural diagram of an intelligent operation and maintenance management system for new energy equipment. Figure 2 This is a flowchart of the fault risk prediction method; Figure 3 It is a flowchart of the interactive display and adaptive calibration process for the operation and maintenance of new energy power stations. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] See Figure 1 This invention provides an intelligent operation and maintenance management system for new energy equipment, including a data acquisition unit that collects multi-source heterogeneous operation and maintenance data from distributed new energy equipment and fuses the data according to equipment identifiers and timestamps to generate a standardized operation and maintenance dataset. A status assessment unit, based on the standardized operation and maintenance dataset, uses a pre-trained graph neural network to model the spatiotemporal coupling relationship between new energy equipment, obtaining an equipment health status map. A risk prediction unit, based on the equipment health status map, uses a temporal convolutional network to extrapolate future operating trends of the equipment, generating a fault risk prediction result with an uncertainty interval. A maintenance decision unit receives the fault risk prediction result and, through a reinforcement learning agent, performs multi-objective optimization between operation and maintenance costs and power generation revenue, outputting equipment maintenance strategy suggestions. An interactive display unit transforms the equipment maintenance strategy suggestions into a three-dimensional dynamic visualization interface and responds to operation commands from maintenance personnel, adjusting the equipment parameters in the three-dimensional dynamic visualization interface accordingly.

[0024] Example 1: In practice, the data acquisition unit collects multi-source heterogeneous operation and maintenance data from distributed renewable energy equipment. This equipment includes wind turbines, photovoltaic modules, and energy storage batteries. The data acquisition unit obtains vibration data, current and voltage data, and temperature data from the wind turbines via sensor interfaces; it also obtains current and voltage data and temperature data from the photovoltaic modules; and current and voltage data and temperature data from the energy storage batteries. Furthermore, the data acquisition unit acquires meteorological environmental data, including wind speed, irradiance, ambient temperature, and ambient humidity, through a meteorological monitoring station interface.

[0025] The data acquisition unit correlates and fuses multi-source heterogeneous operation and maintenance data. It groups wind turbine vibration data, current and voltage data, temperature data, and meteorological data according to the corresponding equipment identifier, forming wind turbine data groups. Similarly, it groups photovoltaic module current and voltage data, temperature data, and meteorological data according to the corresponding equipment identifier, forming photovoltaic data groups. Likewise, it groups energy storage battery current and voltage data, temperature data, and meteorological data according to the corresponding equipment identifier, forming energy storage data groups. Within each group, the data acquisition unit aligns the acquisition times of all data points to a unified time grid, using timestamps as a reference. For data points with time discrepancies, linear interpolation is used to generate complete data vectors at the same timestamp within each group. Finally, the data acquisition unit converts the aligned and interpolated multi-source data into a standardized field sequence in a unified format. The unified format includes fields for equipment identification, timestamp, vibration amplitude, vibration frequency, current, voltage, temperature, wind speed, irradiance, ambient temperature, and ambient humidity. All fields are converted into standard data types according to preset numerical encoding rules to form a standardized operation and maintenance dataset.

[0026] The condition assessment unit obtains the operating parameter sequence for each device in the standardized operation and maintenance dataset. The operating parameter sequence is a multi-dimensional time series formed by the standardized field sequence at consecutive timestamps for each device extracted from the standardized operation and maintenance dataset. Each time step vector includes vibration amplitude, vibration frequency, current value, voltage value, temperature value, wind speed, irradiance, ambient temperature, and ambient humidity. The condition assessment unit constructs an initial graph structure, with each device as a node and the electrical connections between devices as edges. The electrical connections are determined based on the electrical single-line diagram of the new energy power station. When two devices are directly connected via cables or busbars, there is an edge between the corresponding two nodes. The initial graph structure is represented as G=(V,E), where V represents the set of nodes and E represents the set of edges.

[0027] The state evaluation unit uses a pre-trained graph neural network to model the spatiotemporal coupling relationship between new energy devices. The pre-trained graph neural network adopts a graph attention network architecture, which includes three multi-head attention graph convolutional layers and one fully connected output layer. The first multi-head attention graph convolutional layer maps the node feature dimension from the input dimension to 128 dimensions; the second multi-head attention graph convolutional layer maintains the 128-dimensional features; and the third multi-head attention graph convolutional layer maps the feature dimension to 64 dimensions. Each multi-head attention graph convolutional layer contains four attention heads, and each attention head calculates the attention coefficient between node i and node j. The formula for calculating the attention coefficient is: ; in, Represents a node The input feature vector, Represents a node The input feature vector, This represents the learnable weight matrix. This represents a learnable attention weight vector. This indicates the transpose operation. This represents a vector concatenation operation. Represents a node The set of neighboring nodes is defined by LeakyReLU, which represents the Leaky Rectified Linear Activation Function. Layer normalization and exponential linear unit activation are applied after each multi-head attention map convolutional layer. The fully connected output layer maps the 64-dimensional features to the dimensions of a health score vector, which contains three components: the overall device health index, the critical component degradation index, and the remaining effective lifespan index.

[0028] The training process of the pre-trained graph neural network is as follows: Historical operation and maintenance datasets of new energy power plants are acquired. These datasets contain standardized operation and maintenance data spanning at least three years and corresponding actual equipment status labels. The actual equipment status labels include fault types, component replacement records, and manual inspection scores from maintenance records. Multiple graph structure samples are constructed from the historical standardized operation and maintenance data according to equipment identification and time sequence. Each graph structure sample corresponds to a time window with a length of 24 hours and a sliding step of 1 hour. Node features are vectors derived from the average pooling of the sequence of operating parameters for each device within the time window. Edge connections are consistent with the electrical connections in the initial graph structure. The training objective of the graph neural network is to minimize the mean squared error loss between the predicted value of the health score vector and the actual equipment status label. The Adam optimizer is used for parameter updates, with an initial learning rate of 0.001, which decays to 0.0001 after 60 training iterations. After training, a pre-trained graph neural network is obtained. The model parameters of the pre-trained graph neural network are then solidified and deployed to the state evaluation unit.

[0029] When the state evaluation unit invokes the pre-trained graph neural network, it inputs the operating parameter sequence of each device as node features into the pre-trained graph neural network. The operating parameter sequence is processed as follows: for each device, the standardized field sequence from the most recent 24 hours is input into a one-dimensional convolutional encoder in chronological order. The one-dimensional convolutional encoder contains two dilated convolutional layers, each with a kernel size of 3, dilation factors of 1 and 2 respectively, and 32 output channels. The one-dimensional convolutional encoder encodes the operating parameter sequence of each device into a fixed-length node feature vector. The state evaluation unit inputs the node feature vectors of each device and the initial graph structure into the pre-trained graph neural network. The pre-trained graph neural network propagates node features through three multi-head attention graph convolutional layers and aggregates the spatiotemporal dependency features of neighboring nodes. The spatiotemporal dependency features capture the mutual influence between devices at adjacent time steps through a multi-layer attention mechanism. After the three-layer graph convolution operation, the output features of each node are input into the fully connected output layer, mapped to a health score vector. The status assessment unit binds the health score vector to the corresponding node in the initial graph structure to generate a device health status graph. The device health status graph is a graph data structure that includes node locations, edge connections, and the health score vector of each node.

[0030] In some embodiments, before extracting node feature vectors, the one-dimensional convolutional encoder normalizes the sequence of operating parameters using Z-score standardization. The mean and standard deviation of each field are calculated using the statistical values ​​of the full dataset from the training phase of the pre-trained graph neural network. After the node feature vectors are generated, they are input into the pre-trained graph neural network along with the initial graph structure for forward propagation. During forward propagation, the node features output from each multi-head attention map convolutional layer are residually connected to the input features to prevent gradient vanishing in deep networks. The overall device health index in the health score vector ranges from 0 to 1, with higher values ​​indicating healthier device conditions. The critical component degradation index ranges from 0 to 1, with higher values ​​indicating more severe degradation of critical components. The remaining effective lifespan index is in hours, representing the expected stable operating time of the device under the current operating conditions. The device health status map is stored in JSON format, which includes node identifiers, node coordinates, values ​​of each component of the health score vector, and an edge adjacency list.

[0031] Example 2: In specific implementation, please refer to Figure 2The risk prediction unit extracts the health score sequence for each device from the equipment health status map. Each node in the equipment health status map is bound to a health score vector, which includes the overall equipment health index, the critical component degradation index, and the remaining effective life index. The risk prediction unit retrieves the health score vectors of the same device from the equipment health status map across multiple consecutive time steps, unfolding each time step's health score vector into one-dimensional sequence elements to form the time series vector for each device. The risk prediction unit then arranges the three components of each device's time series vector along the time axis to obtain three basic health score sequences for each device: the overall equipment health index sequence, the critical component degradation index sequence, and the remaining effective life index sequence.

[0032] The risk prediction unit also extracts the association strength sequence between adjacent devices from the device health status map. Adjacent devices are determined by edges in the device health status map. The risk prediction unit calculates the health score similarity of the two devices connected by each edge at each time step. The health score similarity is calculated using cosine similarity, which takes the health score vectors of the two devices as input and outputs a scalar value, which serves as the association strength. For each edge, the risk prediction unit arranges the association strengths sequentially along the time steps, forming the association strength sequence corresponding to each edge.

[0033] The risk prediction unit concatenates the overall health index sequence, critical component degradation index sequence, remaining effective life index sequence, and correlation strength sequence of each edge between the target device and each adjacent device for each target device. The concatenation method is as follows: at each time step, a feature vector is formed by combining the three health scores of the target device and all correlation strength values ​​between the target device and all adjacent devices. The feature vectors of all time steps are arranged in chronological order to form a multi-channel time-series input matrix. The number of rows in the multi-channel time-series input matrix equals the number of time steps, and the number of columns equals 3 plus the number of adjacent devices of the target device.

[0034] In some embodiments, when the number of neighboring devices of the target device is less than the preset maximum number of neighbors, the risk prediction unit fills the missing association strength channels with zero values; when the number of neighboring devices of the target device exceeds the preset maximum number of neighbors, the risk prediction unit selects the association strength channels corresponding to the first few edges with the largest absolute value of association strength, and the maximum number of neighbors is set to 10.

[0035] The risk prediction unit inputs a multi-channel temporal input matrix into a temporal convolutional network. The core architecture of the temporal convolutional network consists of multiple stacked residual modules. Each residual module contains two dilated convolutional layers, a weight normalization operation, a linear rectified activation function, and a spatial random deactivation operation. A one-dimensional convolutional mapping residual connection is added between the input and output of the residual module.

[0036] In practical implementation, the input layer of the temporal convolutional network receives a multi-channel temporal input matrix with the shape [T, C_in], where T represents the number of historical time steps, C_in represents the number of input channels, and C_in equals 3 plus the number of effective adjacent devices. The temporal convolutional network contains four residual modules with dilation factors set to 1, 2, 4, and 8, respectively. The first dilated convolutional layer within each residual module uses a kernel size of 3, a stride of 1, and a dilation factor corresponding to the module; the number of kernels is uniformly set to 64. The second dilated convolutional layer uses a kernel size of 3, a stride of 1, and the same dilation factor as the first dilated convolutional layer; the number of kernels is also uniformly set to 64. The output calculation formula for the first dilated convolutional layer is: ; in, Indicates the first The input matrix of the layer residual module at time step and the The value on each channel Indicates the first The expansion factor of the layer residual module, Indicates the first The expansion factor in the layer residual module is The first convolution kernel in the dilated convolution layer The position, the The input channel to the first The weight parameters of each output channel Indicates the first The expansion factor in the layer residual module is The first dilated convolutional layer The bias parameters of each output channel. Indicates the first The dilated convolutional layer of the residual module at time step and the The output value on each output channel Indicates the first Number of input channels for the layer residual module.

[0037] Each residual module performs two dilated convolutions on the input, and then adds the residual connections (through one-dimensional convolutional mappings) to the module input to form the output of the residual module. After processing by all residual modules, the temporal convolutional network outputs the prediction result through a fully connected output layer. The fully connected output layer contains two parallel linear projectors. The first linear projector outputs the predicted mean of the health score at the next prediction time, and the second linear projector outputs the logarithmic value of the predicted variance of the health score at the next prediction time. The output dimension of the prediction mean is [F, 3], and the output dimension of the logarithmic value of the prediction variance is [F, 3], where F represents the number of future prediction times, and 3 corresponds to the three components of the health score vector.

[0038] The training process of the temporal convolutional network is as follows: Training samples are constructed using a historical standardized operation and maintenance dataset and a sequence of historical equipment health status maps. The input to each training sample is a multi-channel temporal input matrix of length T for historical time steps, and the training labels are the sequence of true health score vectors for the corresponding future F time steps. The true health score vector sequence is generated based on subsequent equipment maintenance records and status assessment results. During training, the temporal convolutional network updates its parameters by minimizing the negative log-likelihood loss function. The negative log-likelihood loss function combines the predicted mean and the predicted variance, assuming that each component of the future health score vector follows a Gaussian distribution. The optimizer uses Adam, with an initial learning rate of 0.0005, a batch size of 32, and 500 training epochs. After training, the weight parameters of the temporal convolutional network are fixed and deployed to the risk prediction unit.

[0039] The risk prediction unit calculates the health score prediction interval for multiple future prediction times using dilated convolutional layers and residual connection layers in a temporal convolutional network. The number of future prediction times is set to 24, corresponding to one prediction time per hour over the next 24 hours. For each future prediction time, the temporal convolutional network outputs a predicted mean vector and a predicted variance vector. The risk prediction unit constructs a health score prediction interval centered on the predicted mean and with a radius of twice the predicted standard deviation. Specifically, the lower bound of the interval is the predicted mean minus twice the predicted standard deviation, and the upper bound is the predicted mean plus twice the predicted standard deviation. The predicted standard deviation is obtained by exponentially calculating the logarithm of the predicted variance.

[0040] The risk prediction unit compares the predicted interval of the health score with preset thresholds. The preset thresholds are set for the three components of the health score vector: the overall equipment health index threshold is set to 0.3; when the upper bound of the overall equipment health index prediction interval is below 0.3, the component is considered to be at risk. The critical component deterioration index threshold is set to 0.7; when the lower bound of the critical component deterioration index prediction interval is above 0.7, the component is considered to be at risk. The remaining effective lifespan index threshold is set to 720 hours; when the upper bound of the remaining effective lifespan index prediction interval is below 720 hours, the component is considered to be at risk. When the prediction interval of any health score component meets the corresponding threshold condition, the risk prediction unit marks the corresponding future prediction time as a risk time.

[0041] The risk prediction unit combines the risk time, the corresponding equipment identifier, and the corresponding health score prediction interval data to generate a fault risk prediction result. The fault risk prediction result is output in the form of a structured data list, with each record containing the risk equipment identifier, the risk time, the name of the health score component that triggered the risk, and the corresponding prediction interval boundary value.

[0042] Example 3: In practice, the maintenance decision-making unit obtains the risk moments and corresponding risk equipment identifiers marked in the fault risk prediction results. The fault risk prediction results are input into the maintenance decision-making unit in the form of a structured data list. Each record in the data list includes a risk equipment identifier field, a risk moment field, a health score component name field that triggers the risk, and a corresponding prediction interval boundary value field. The maintenance decision-making unit parses the data list, extracts all risk equipment identifiers and all risk moments associated with each risk equipment identifier, and organizes the extracted results into a risk equipment mapping table. The risk equipment mapping table uses the risk equipment identifier as the key and the list of risk moments as the value.

[0043] The maintenance decision-making unit obtains the current inventory of maintenance resources, which is retrieved in real-time from the maintenance resource management database of the new energy power station. The maintenance resource management database stores inventory data including: spare parts inventory quantity, available maintenance tool quantity, available man-hours for maintenance teams, and available transport vehicle quantity. Each inventory data point is timestamped, allowing the maintenance decision-making unit to obtain the latest inventory value at the time of the query.

[0044] The maintenance decision-making unit also obtains power generation forecast curves for future periods. These curves are provided by the power prediction system of the renewable energy power station, which forecasts future power generation based on numerical weather prediction data and historical power generation data. The power generation forecast curves have a time resolution of one hour, covering a time span of the next 72 hours. Each hour corresponds to a predicted power generation value, measured in kilowatts. The power generation forecast curves are input to the maintenance decision-making unit in array format, with an array length of 72.

[0045] The maintenance decision-making unit combines the risk equipment identifier, the current inventory of maintenance resources, and the power generation forecast curve for future periods into a state space vector. The encoding method for the state space vector is as follows: risk equipment identifiers use unique hot coding, with the length of the unique hot coding being the total number of equipment in the new energy power station. The risk equipment identifier's corresponding position is set to 1, and the remaining positions are set to 0. The current inventory of maintenance resources uses numerical coding, normalizing the inventory quantity of spare parts, the available quantity of maintenance tools, the available working hours of maintenance teams, and the available quantity of transport vehicles to the 0-1 range. The normalization calculation method is to divide the numerical value of each item's inventory by the historical maximum value of that item's inventory. The power generation forecast curve uses 72 normalized power generation values ​​arranged sequentially. Each normalized power generation value is calculated by dividing the corresponding predicted power generation value by the rated installed capacity of the new energy power station. The total dimension of the state space vector is: the total number of equipment plus the 4-item maintenance resource inventory coding dimension and the 72-dimensional power generation forecast coding dimension.

[0046] The core architecture of the reinforcement learning agent employs a deep Q-network. The deep Q-network consists of an online Q-network and a target Q-network, with identical network structures. The online Q-network takes a state space vector as its input and its structure is as follows: a first fully connected layer with an input dimension equal to the total dimension of the state space vector and an output dimension of 256; a second fully connected layer with both input and output dimensions of 256; a third fully connected layer with an input dimension of 256 and an output dimension of 128; and a fourth fully connected layer with an input dimension of 128 and an output dimension equal to the action space dimension. A linear rectified activation function is applied after each fully connected layer, but no activation function is applied to the output layer of the online Q-network. The target Q-network has the same network structure as the online Q-network, and its parameters are periodically copied from the online Q-network during training.

[0047] The action space consists of maintenance actions. Each maintenance action comprises a combination of three dimensions: maintenance time selection, maintenance type selection, and maintenance resource allocation selection. The maintenance time selection dimension is discretized into six options: immediate execution, execution with a 1-hour delay, execution with a 2-hour delay, execution with a 4-hour delay, execution with an 8-hour delay, and execution with a 24-hour delay. The maintenance type selection dimension is discretized into four options: routine inspection, component replacement, in-depth maintenance, and online repair without downtime. The maintenance resource allocation selection dimension is discretized into three options: minimum resource allocation, standard resource allocation, and maximum resource allocation. The combinations of these three dimensions form the complete action space, which contains a total of 6 x 4 x 3 = 72 discrete actions. The output dimension of the online Q-network corresponds to these 72 actions, outputting the action value Q-value for each action.

[0048] The training process of the reinforcement learning agent employs an experience replay mechanism. Before training begins, an experience replay buffer is constructed, with a capacity of 100,000 experience samples. Each experience sample contains the current state space vector, the selected maintenance action number, the reward value obtained after executing the maintenance action, the next state space vector after executing the maintenance action, and an indicator indicating whether the termination state has been reached. The specific training steps are as follows: Initialize the parameters of the online Q-network to random values, and initialize the parameters of the target Q-network to be the same as those of the online Q-network; for each round of training, extract a state sequence for a continuous time period from the historical operation and maintenance data of the new energy power station. The state sequence contains state space vectors for multiple consecutive time steps; at each time step, randomly select a maintenance action from the action space with probability ε, and select the maintenance action corresponding to the maximum Q value output by the online Q-network with probability 1-ε. ε decreases linearly from 1.0 to 0.01, with a decay step count of 50,000 steps; after executing the selected maintenance action, obtain the reward value from the simulation environment and transform it into the next state space vector. The simulation environment is constructed based on historical actual operation and maintenance results and power generation revenue data; store the single-step experience samples in the experience replay buffer. When the number of samples in the experience replay buffer reaches 1000, start updating the network parameters; for each parameter update, randomly sample 64 experience samples from the experience replay buffer and calculate the target value. The formula for calculating the target value is: ; in, Indicates the first The target value of the sampled empirical data. Indicates the first The reward value of each sample of experience. Indicates the discount factor. Set to 0.99, Indicates the first The next state space vector of the sampled empirical samples For all possible actions in the target Q network The maximum action value Q is output. This represents the set of parameters for the target Q-network.

[0049] The online Q-network updates its parameters by minimizing the mean squared error between the predicted Q-value and the target value. The optimizer used is Adam, and the learning rate is set to 0.0001. After every 2000 steps of parameter updates, the parameters of the online Q-network are copied to the target Q-network. Training continues until the cumulative reward function converges. Training stops when the cumulative reward function does not show an improvement of more than 1% on the validation set for 50 consecutive epochs.

[0050] The cumulative reward function is designed to be negatively correlated with operation and maintenance costs and positively correlated with power generation revenue. The specific components of the cumulative reward function are as follows: The reward function includes a basic reward item, a maintenance cost penalty item, and a risk penalty item. The basic reward item equals the ratio of actual power generation revenue in the future period to the theoretical power generation revenue under fault-free conditions multiplied by a fixed benchmark revenue value, which is set to 1000. The maintenance cost penalty item is calculated based on the resource consumption cost corresponding to the maintenance resource allocation amount. The resource consumption cost includes the sum of the value of spare parts consumption, maintenance labor costs, and transportation costs, divided by a benchmark resource cost coefficient, which is set to 5000. The risk penalty item is triggered if the equipment still fails after the maintenance action is performed. The risk penalty item takes a fixed penalty value, which is set to -10000.

[0051] During the inference phase after training, the maintenance decision unit inputs the state space vector into the online Q-network, which outputs the Q-values ​​of 72 actions. The maintenance decision unit selects the action with the highest Q-value as the optimal maintenance action. The optimal maintenance action decodes the specific values ​​of maintenance time, maintenance type, and maintenance resource allocation, forming a suggested equipment maintenance strategy. The suggested equipment maintenance strategy is output in structured data format, including risk equipment identification, suggested maintenance time, suggested maintenance type, suggested maintenance resource allocation, a list of allocation quantities for each resource, and estimated maintenance time.

[0052] Example 4: In specific implementation, please refer to Figure 3 The interactive display unit receives equipment maintenance strategy suggestions from the maintenance decision-making unit. These suggestions include the identifier of each risky device, the suggested maintenance time, the suggested maintenance type, the suggested maintenance resource allocation, and a list of allocation quantities for each resource. Based on the device identifier in the maintenance strategy suggestions, the interactive display unit loads the corresponding 3D geographic coordinates of the new energy power station from its 3D geographic information model database. The 3D geographic information model database stores the longitude, latitude, altitude, and orientation angle of each device in a unified geographic coordinate system.

[0053] The interactive display unit generates a corresponding 3D model in a 3D geographic coordinate system for each device in the equipment maintenance strategy recommendations. The 3D model generation method involves: calling a pre-made 3D geometric model corresponding to the device type from the equipment 3D model library; translating the pre-made 3D geometric model according to the device's 3D geographic coordinates; rotating it according to the device's orientation angle; and adapting it to a uniform scaling factor to generate a 3D model instance placed in the 3D geographic coordinate system. The surface material of the 3D model instance is texture-mapped based on the actual appearance image of the device.

[0054] The interactive display unit overlays a health score vector from the device's health status graph onto the 3D model. The health score vector comprises three components: overall device health index, critical component degradation index, and remaining effective lifespan index. These three components are attached to the 3D model as semi-transparent colored labels, fixed at the top center of the model's bounding box. The overall device health index is displayed as a green-to-red gradient bar, with green corresponding to a value of 1.0 and red to a value of 0.0. The critical component degradation index is displayed as a yellow-to-dark red gradient bar, with yellow corresponding to a value of 0.0 and dark red to a value of 1.0. The remaining effective lifespan index is displayed as a combination of numerical text and the unit "hours."

[0055] The interactive display unit receives click operations from maintenance personnel on the 3D models. Click operation detection is achieved through the ray casting algorithm of the 3D graphics engine. The interactive display unit emits a ray from the 2D coordinates of the mouse cursor on the screen, calculates the intersection of the ray with the bounding boxes of all 3D models, and identifies the device corresponding to the first intersecting 3D model as the selected device. The interactive display unit also receives bounding box operations from maintenance personnel. Bounding selection operations are performed by drawing a rectangular selection box on the screen to select all 3D models within the rectangular selection box. The outer contour of the selected 3D model is highlighted with a bright blue outline, and the outline width is two screen pixels.

[0056] Upon receiving a click or selection action, the interactive display unit retrieves the detailed operating parameters corresponding to the selected device from the standardized operation and maintenance dataset. This standardized dataset is stored on a data management server. The interactive display unit sends a query request to the data management server using the device identifier, which includes the device identifier and the start and end timestamps of the current display time window. The data management server returns a matching standardized field sequence, including complete time-series data for vibration amplitude, vibration frequency, current, voltage, temperature, wind speed, irradiance, ambient temperature, and ambient humidity. The interactive display unit creates a floating window in the 3D view, which floats to the right of the selected device as a semi-transparent panel. The floating window displays detailed operating parameters in a combination of lists and graphs. The list displays the current values ​​of each field row by row, while the graph shows the trend of each field over the past 24 hours.

[0057] The interactive display unit processes parameter adjustment commands input by maintenance personnel. These commands can be input via: personnel directly modifying parameter values ​​in the input boxes within the floating window, or adjusting equipment operating parameters by dragging virtual control handles on the 3D model. Upon receiving the command, the interactive display unit synchronously updates the 3D model's display status, including updating the health score vector label values ​​overlaid on the model and updating the color mapping. Simultaneously, the interactive display unit triggers the status assessment unit to recalculate the health scores of associated devices. This is achieved by combining the adjusted parameters into a parameter update message, which includes the device identifier of the adjusted device, the adjusted parameter field names, and the adjusted parameter values, and then sending this message to the status assessment unit. Upon receiving the update message, the status assessment unit extracts the operating parameter sequences of associated devices from the standardized maintenance dataset. Associated devices are those with edge connections to the adjusted device in the device health status graph. The status assessment unit then re-executes the graph neural network forward propagation to generate updated health score vectors for the associated devices and returns these vectors to the interactive display unit for display refresh.

[0058] The interactive display unit also supports maintenance personnel in performing sectioning operations on the 3D model. The sectioning operation is activated by the maintenance personnel clicking the sectioning tool button in the interactive display unit's tool panel, and then dragging on the 3D model to define the sectioning plane. The orientation of the sectioning plane is determined by the dragging direction. Based on the spatial relationship between the sectioning plane and the 3D model, the interactive display unit calculates the clipping plane of the 3D model, renders the portion of the 3D model inside the clipping plane, hides the portion outside the clipping plane, and displays the internal structure of the equipment at the sectioning plane. The 3D representation of the internal structure is loaded from a library of key internal component models, including the wind turbine gearbox, generator stator and rotor, photovoltaic module junction box and bypass diodes, energy storage battery cell modules, and thermal management system piping. At the sectioning plane, the interactive display unit overlays and displays real-time temperature field and stress distribution cloud maps of the key internal components. Real-time temperature field data is collected by a distributed temperature sensor network. Temperature sensors are installed on the surfaces of key components inside the equipment, with a sampling frequency of once per minute. The temperature field data is stored in point cloud format. The interactive display unit uses a spatial interpolation algorithm to convert the point cloud temperature values ​​into continuous color block textures, which are then rendered as a heatmap on the cross-sectional surface. Stress distribution cloud map data is calculated online by the finite element analysis model based on the equipment's real-time operating loads and structural parameters. The stress distribution cloud map is expressed using contour lines and color fills.

[0059] The adaptive calibration unit receives maintenance operation records actually performed by maintenance personnel from the interactive display unit. These records are automatically generated and pushed to the adaptive calibration unit by the site maintenance management system after a maintenance work order is completed. The maintenance operation records include fields for equipment identification, actual maintenance time, actual maintenance type, actual maintenance resource consumption, and actual maintenance result remarks. The adaptive calibration unit parses the maintenance operation records, extracting the actual maintenance time and actual maintenance resource consumption corresponding to each equipment identification. The actual maintenance resource consumption is further subdivided into the actual quantity of spare parts consumed, the actual quantity of repair man-hours consumed, and the actual number of times transport vehicles were used.

[0060] The adaptive calibration unit extracts the recommended maintenance time and recommended maintenance resource consumption corresponding to the same equipment identifier from the equipment maintenance strategy recommendations. The recommended maintenance resource consumption is further subdivided into the recommended consumption quantity of spare parts, the recommended consumption quantity of maintenance man-hours, and the recommended number of times the transport vehicle is used.

[0061] The adaptive calibration unit combines the time difference between the actual maintenance time and the recommended maintenance time, and the resource difference between the actual maintenance resource consumption and the recommended maintenance resource consumption, into a deviation vector. The deviation vector is constructed as follows: the time difference is converted to a real number in hours, the difference in spare parts consumption is converted to an integer, the difference in maintenance man-hour consumption is converted to a real number in hours, and the difference in the number of times the transport vehicle is used is converted to an integer. The four differences are arranged in order to form a four-dimensional deviation vector.

[0062] The adaptive calibration unit uses the deviation vector to perform online gradient updates on the node weights of the graph neural network and the reward function parameters of the reinforcement learning agent. Specifically, the online gradient update method is as follows: the adaptive calibration unit obtains the actual health status label of the risk equipment identifier after the actual maintenance operation from the operation and maintenance management system. This actual health status label includes the health score vector of the maintained equipment (overall equipment health index, key component degradation index, and remaining effective lifespan index), determined by manual inspection records or equipment shutdown detection results. The adaptive calibration unit uses the deviation vector as a supervision signal to construct a comprehensive loss function. This comprehensive loss function is obtained by summing three parts: the prediction loss term of the graph neural network output layer (…). ), calculate the mean squared error between the health score vector output by the graph neural network and the actual health status label after maintenance; the value loss term of the reinforcement learning agent policy network ( The sum of the absolute values ​​of each component of the bias vector is calculated; the elastic weights reinforce the regularization term. The mathematical expression of the comprehensive loss function is: ; in, This represents the value of the comprehensive loss function; This represents the prediction loss term of the graph neural network output layer, calculated as the mean square error between the health score vector output by the graph neural network and the actual health status label after maintenance. This represents the value loss term of the policy network of the reinforcement learning agent, and is calculated as the sum of the absolute values ​​of each component of the bias vector; This represents the strength coefficient of the elastic weight consolidation regularization term, with a value of 0.01; Indicates the first Model parameters The estimated values ​​of the diagonal elements of the corresponding Fisher information matrix are calculated from the empirical Fisher information of historical operation and maintenance data; Indicates the number before the current update. Each model parameter value; Indicates the first The optimal parameter values ​​of each model after the previous calibration.

[0063] Calculate the diagonal element estimates of the Fisher information matrix The method is as follows: The adaptive calibration unit stores a calibration dataset formed by the most recent 100 maintenance operation records. The graph neural network and reinforcement learning agent then compute their respective outputs for each parameter on this calibration dataset. The expected value of the squared gradient is used as... The role of the elastic weight consolidation regularization term is to suppress the catastrophic forgetting of historical key features by node weights and reward function parameters during the update process. This is achieved by applying deviation penalties to important parameters, which are determined by... Logo, The larger the value, the more important the parameter is to the historical task.

[0064] The adaptive calibration unit calculates the first loss gradient of the output layer of the graph neural network. The first loss gradient is... ,in This represents the set of weights for all trainable nodes in the graph neural network. The adaptive calibration unit computes the second loss gradient of the reinforcement learning agent policy network, which is: ,in This represents the set of reward function parameters and Q-network parameters in the reinforcement learning agent. The adaptive calibration unit updates the node weights of the graph neural network using the Adam optimizer based on the first loss gradient, with a learning rate set to 0.0001; based on the second loss gradient, it updates the reward function parameters and Q-network parameters of the reinforcement learning agent using the Adam optimizer, with a learning rate set to 0.0001. After the updates are complete, a calibrated graph neural network and a calibrated reinforcement learning agent are generated.

[0065] The adaptive calibration unit serializes the complete model parameters of the calibrated graph neural network and sends them to the state evaluation unit via the model distribution interface. Upon receiving the parameters, the state evaluation unit replaces the original graph neural network model parameters, and subsequent health status evaluation tasks are performed using the calibrated graph neural network. Similarly, the adaptive calibration unit serializes the complete model parameters of the calibrated reinforcement learning agent and sends them to the maintenance decision unit via the model distribution interface. Upon receiving the parameters, the maintenance decision unit replaces the original reinforcement learning agent model parameters, and subsequent maintenance policy generation tasks are performed using the calibrated reinforcement learning agent.

[0066] In some embodiments, the online gradient update process of the adaptive calibration unit is completed within 30 seconds after each maintenance operation record feedback to ensure that the system model parameters can adapt to the actual operational deviations of maintenance personnel in a timely manner. Elastic weights reinforce the strength coefficient of the regularization term. The value is determined based on the cross-validation results that balance the performance of the old and new tasks in the offline validation set. Setting it to 0.01 allows for effective fitting of the deviation vector while maintaining the accuracy loss of historical tasks to no more than 3%.

[0067] Example 5: In practical implementation, after the data acquisition unit generates the standardized operation and maintenance dataset, the data integrity verification unit performs integrity checks on the standardized operation and maintenance dataset. The data integrity verification unit receives the complete file path of the standardized operation and maintenance dataset, parses each data record in the dataset line by line, and extracts the device identifier field and timestamp field from each data record. The data integrity verification unit groups the extracted device identifiers and timestamps according to the device identifier, constructing a timestamp sequence set corresponding to each device identifier. Based on a preset sampling time step, the data integrity verification unit checks whether there are any missing data intervals in the timestamp sequence corresponding to each device identifier. The preset sampling time step is determined by the acquisition frequency of the data acquisition unit, which is once every 10 minutes; therefore, the preset sampling time step is 10 minutes. The data integrity verification unit arranges the timestamp sequence in ascending order and calculates the time interval between adjacent timestamps. When the time interval between adjacent timestamps exceeds 1.5 times the sampling time step, i.e. 15 minutes, the time period between the two timestamps is marked as a data missing interval. The start time of the data missing interval is the previous timestamp plus the sampling time step, and the end time is the next timestamp minus the sampling time step.

[0068] For device identifiers with missing data ranges, the data integrity verification unit searches for other device identifiers spatially adjacent to this device identifier. Spatial adjacency is determined based on the geographical coordinates of the new energy power station equipment. The data integrity verification unit reads the longitude, latitude, and altitude coordinates of each device from the power station equipment information database. The data integrity verification unit calculates the spatial Euclidean distance between the missing device and other devices of the same type. When the calculated spatial Euclidean distance is less than a preset spatial adjacency distance threshold, the corresponding other device identifier is marked as a spatially adjacent device identifier. The spatial adjacency distance threshold is set to 500 meters, based on the statistical value of the average interval distance of the equipment cluster within the new energy power station. The average interval distance is calculated from the equipment coordinates on the power station's overall layout plan. 500 meters can cover the main related equipment within the same power station area.

[0069] The data integrity verification unit obtains the data for each spatially adjacent device identifier at timestamps within the same missing data interval. Specifically, for each missing timestamp within the missing data interval, it queries the standardized field sequence of each spatially adjacent device identifier at the same missing timestamp from the standardized operation and maintenance dataset. The data integrity verification unit then fills in the missing data interval using spatial interpolation. The spatial interpolation method employs an inverse distance weighted interpolation algorithm, and the filling calculation formula is as follows: ; in, Indicates device identification Missing timestamp The generated normalized field sequence vector is filled in. For devices with missing data ranges, For missing timestamps within the data missing range, Indicates device identification The total number of spatially adjacent device identifiers. An index for identifying spatially adjacent devices. The value range is 1 to integers, Indicates the first The spatially adjacent device identifiers are in the missing timestamp. The standardized field sequence vector at the location, Indicates device identification With the The spatial distance between adjacent device identifiers is calculated by taking the square root of the sum of the squares of the differences in longitude, latitude, and altitude coordinates between the two devices. This represents the power-law exponent of distance decay. The setting is 2, based on the fact that in the inverse distance weighted interpolation algorithm, the squared distance decay can avoid excessive smoothing of data from distant devices while maintaining computational efficiency, so that the interpolation results are mainly affected by nearby devices.

[0070] The data integrity verification unit fills in the generated data and writes it into the corresponding missing positions in the standardized operation and maintenance dataset, forming a complete standardized operation and maintenance dataset. The timestamp sequences of all device identifiers in the complete standardized operation and maintenance dataset meet the requirement of sampling time step continuity. The data integrity verification unit then sends the complete standardized operation and maintenance dataset to the status assessment unit via a message queue. The message queue uses a reliable transmission protocol to ensure data integrity.

[0071] The confidence assessment unit obtains the uncertainty interval from the fault risk prediction results generated by the risk prediction unit. Each record in the fault risk prediction results contains three parts: the risky device identifier, the risk time, and the health score prediction interval. The health score prediction interval provides a lower bound and an upper bound for each of the three components of the health score vector. The confidence assessment unit extracts the lower bound and upper bound of the interval for each of the three components of each risky device identifier at each risk time, and calculates the uncertainty interval width for each component. The uncertainty interval width is calculated by subtracting the lower bound from the upper bound.

[0072] The confidence assessment unit compares the uncertainty interval width of each component with a preset confidence threshold. The preset confidence thresholds are set for the three components of the health score vector: 0.2 for the overall equipment health index, 0.2 for the critical component deterioration index, and 500 hours for the remaining effective lifespan index. The three confidence thresholds are set based on the following: according to historical accuracy analysis of new energy equipment fault prediction, when the uncertainty interval width of the health score component is lower than the corresponding threshold, the reliability of the prediction result reaches over 90%, meeting the conditions for direct early warning; when the uncertainty interval width exceeds the corresponding threshold, the prediction result has a large random bias.

[0073] When the uncertainty interval width of any component exceeds the corresponding confidence threshold, the confidence assessment unit determines that the data sampling frequency for the corresponding device needs to be increased. The confidence assessment unit generates a sampling frequency adjustment instruction, which includes the target device identifier, the new sampling frequency multiplier, and the duration. The new sampling frequency multiplier is set to 5 times, increasing the data acquisition unit's original sampling frequency from once every 10 minutes to once every 2 minutes; the duration is set to the next 24 hours. The confidence assessment unit sends the sampling frequency adjustment instruction to the data acquisition unit, which immediately adjusts the sensor data sampling frequency for the specified device upon receiving the instruction.

[0074] The data acquisition unit collects vibration, current, voltage, and temperature data from the corresponding equipment at an increased sampling frequency. It then processes the newly sampled high-frequency data using the same procedure as generating the standardized maintenance dataset, forming a new, more densely packed sequence of standardized fields. The confidence assessment unit re-inputs the newly sampled high-frequency data into the risk prediction unit. The risk prediction unit re-extracts the health score sequence and correlation strength sequence, constructs a new multi-channel time-series input matrix, performs secondary extrapolation through a time-series convolutional network, outputs an updated health score prediction interval, and generates updated fault risk prediction results.

[0075] The confidence assessment unit recalculates the uncertainty interval width of each component for each risk device identifier at each risk time from the updated failure risk prediction results and compares it with a preset confidence threshold. For risk event records where the uncertainty interval width of all components is lower than the corresponding confidence threshold, the confidence assessment unit marks these risk event records as high-confidence risk events. The data structure of high-confidence risk events includes the risk device identifier, risk time, mean of health score prediction, health score prediction interval, and uncertainty interval width of each component.

[0076] The confidence assessment unit directly pushes high-confidence risk events to the interactive display unit. The push is performed via an internal message bus as a high-priority event notification, with the message format containing the complete data fields of the high-confidence risk event. Upon receiving a high-confidence risk event, the interactive display unit locates the corresponding device's 3D model in the 3D dynamic visualization interface and highlights the 3D model. Specifically, the highlighting operation renders the outer contour of the 3D model as a red outline that flashes twice per second, with the outline width increased to five screen pixels. Simultaneously, a persistent warning information panel is generated next to the 3D model, displaying the risk time, the predicted health score range, and a high-confidence marker icon.

[0077] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A smart operation and maintenance management system for new energy equipment, characterized in that, include: The data acquisition unit is used to collect multi-source heterogeneous operation and maintenance data from distributed new energy equipment, and to associate and fuse the multi-source heterogeneous operation and maintenance data according to the equipment identifier and timestamp to generate a standardized operation and maintenance dataset. The status assessment unit is used to model the spatiotemporal coupling relationship between new energy equipment based on the standardized operation and maintenance dataset and by calling a pre-trained graph neural network to obtain the equipment health status map. The risk prediction unit is used to extrapolate the future operating trend of the equipment based on the equipment health status map and using a temporal convolutional network to generate a fault risk prediction result with an uncertainty interval. The maintenance decision unit is used to receive the fault risk prediction results and perform multi-objective optimization between operation and maintenance costs and power generation revenue through a reinforcement learning agent, and output equipment maintenance strategy suggestions. The interactive display unit is used to construct a three-dimensional dynamic visualization interface based on the equipment maintenance strategy recommendations, and to adjust the equipment parameters in the three-dimensional dynamic visualization interface in response to the operation instructions of the operation and maintenance personnel.

2. The intelligent operation and maintenance management system for new energy equipment according to claim 1, characterized in that, The data acquisition unit associates and fuses the multi-source heterogeneous operation and maintenance data according to device identifiers and timestamps to generate a standardized operation and maintenance dataset in the following specific way: Acquire vibration data, current and voltage data, temperature data, and meteorological environmental data from wind turbines, photovoltaic modules, and energy storage batteries; The vibration data, current and voltage data, temperature data, and meteorological environment data are grouped according to the same device identifier, and aligned and interpolated within each group according to the same timestamp. The multi-source data after alignment and interpolation are converted into a standardized field sequence in a unified format, which serves as the standardized operation and maintenance dataset.

3. The intelligent operation and maintenance management system for new energy equipment according to claim 2, characterized in that, The state assessment unit calls a pre-trained graph neural network to model the spatiotemporal coupling relationship between new energy devices, and obtains the device health status map in the following specific way: Obtain the sequence of operating parameters for each device in the standardized operation and maintenance dataset, and construct an initial graph structure with each device as a node and the electrical connection relationship between devices as an edge. The sequence of operating parameters is input into the graph neural network as node features, and the spatiotemporal dependency features of each device at adjacent time steps are extracted through multi-layer convolution operations of the graph neural network. The spatiotemporal dependent features are mapped to health score vectors, and the health score vectors are bound to the initial graph structure to generate the device health status map.

4. The intelligent operation and maintenance management system for new energy equipment according to claim 3, characterized in that, The risk prediction unit uses a temporal convolutional network to extrapolate the future operating trend of the equipment based on the equipment health status map, and generates a fault risk prediction result with an uncertainty interval in the following specific way: Extract the health score sequence of each device and the correlation strength sequence between adjacent devices from the device health status map; The health score sequence and the correlation strength sequence are concatenated into a multi-channel temporal input matrix, and the multi-channel temporal input matrix is ​​input into the temporal convolutional network. The health score prediction interval for multiple future prediction times is calculated by using the dilated convolutional layer and residual connection layer in the temporal convolutional network, and the time when the health score prediction interval exceeds a preset threshold is marked as a risk time, thereby generating the fault risk prediction result.

5. The intelligent operation and maintenance management system for new energy equipment according to claim 4, characterized in that, The maintenance decision-making unit performs multi-objective optimization between operation and maintenance costs and power generation revenue through a reinforcement learning agent, and outputs equipment maintenance strategy suggestions in the following way: Obtain the marked risk time and corresponding risk device identifier from the fault risk prediction results; The risk equipment identifier, the current inventory of maintenance resources, and the power generation forecast curve for future periods are input into the reinforcement learning agent as the state space; The reinforcement learning agent selects maintenance actions in the action space, which include maintenance time, maintenance type, and maintenance resource allocation. It then outputs the optimal equipment maintenance strategy suggestion by maximizing the cumulative reward function, which is negatively correlated with operation and maintenance costs and positively correlated with power generation revenue.

6. The intelligent operation and maintenance management system for new energy equipment according to claim 5, characterized in that, The interactive display unit constructs a three-dimensional dynamic visualization interface based on the equipment maintenance strategy recommendations, and responds to the operation instructions of maintenance personnel by adjusting the equipment parameters in the three-dimensional dynamic visualization interface in a linked manner. For each device in the equipment maintenance strategy recommendations, a corresponding 3D model is generated in a 3D geographic coordinate system, and the health score vector in the equipment health status map is superimposed and displayed on the 3D model. The system receives clicks or selections from the 3D model by maintenance personnel, retrieves the corresponding detailed operating parameters from the standardized maintenance dataset, and displays them in a floating window. Based on the parameter adjustment instructions input by the maintenance personnel, the display status of the 3D model is updated synchronously, and the status assessment unit is triggered to recalculate the health score of the associated equipment.

7. The intelligent operation and maintenance management system for new energy equipment according to claim 6, characterized in that, The interactive display unit also allows maintenance personnel to perform sectioning operations on the 3D model and display real-time temperature field and stress distribution cloud maps of key components inside the equipment at the sectioning surface.

8. The intelligent operation and maintenance management system for new energy equipment according to claim 6, characterized in that, Also includes: An adaptive calibration unit is used to receive the maintenance operation records actually performed by the operation and maintenance personnel from the interactive display unit, compare the maintenance operation records with the equipment maintenance strategy suggestions, and calculate the deviation vector. The bias vector is used to perform online gradient updates on the node weights of the graph neural network and the reward function parameters of the reinforcement learning agent, thereby generating a calibrated graph neural network and a calibrated reinforcement learning agent. The calibrated graph neural network and the calibrated reinforcement learning agent are sent to the state evaluation unit and the maintenance decision unit, respectively, to replace the original model parameters.

9. The intelligent operation and maintenance management system for new energy equipment according to claim 8, characterized in that, The online gradient update employs an elastic weight consolidation regularization term, which limits the variation of the node weights and the reward function parameters during the update process by applying a deviation penalty to the model parameters.

10. A smart operation and maintenance management system for new energy equipment according to claim 8, characterized in that, The adaptive calibration unit uses the bias vector to perform online gradient updates on the node weights of the graph neural network and the reward function parameters of the reinforcement learning agent as follows: Extract the actual maintenance time and actual maintenance resource consumption from the maintenance operation records, and extract the suggested maintenance time and suggested maintenance resource consumption from the equipment maintenance strategy suggestions; The time difference between the actual maintenance time and the recommended maintenance time, and the resource difference between the actual maintenance resource consumption and the recommended maintenance resource consumption are combined to form the deviation vector; Using the bias vector as a supervision signal, the first loss gradient of the graph neural network output layer is calculated, and the second loss gradient of the reinforcement learning agent Q network is calculated. The node weights and the reward function parameters are updated according to the first loss gradient and the second loss gradient, respectively.

Images