New energy station operation and maintenance system and method based on multi-source data fusion

Abstract

The present application relates to the field of new energy power generation technology, and discloses a new energy station operation and maintenance system and method based on multi-source data fusion. The system comprises a multi-source data acquisition module, an edge data preprocessing module, a feature extraction module, a data fusion engine, a state evaluation module, an operation and maintenance decision optimization module, a digital twin visualization module and a closed-loop feedback module. The method comprises the following steps: collecting multi-source original data; edge cleaning, standardization and spatio-temporal alignment; extracting preliminary fault features; constructing a dynamic confidence fusion model; weighted feature fusion to generate a fusion feature tensor; state evaluation; generating a multi-objective optimization scheduling scheme and executing; closed-loop feedback. By adopting the technical scheme of combining real-time cleaning, spatio-temporal alignment of edge multi-source heterogeneous data and cloud attention mechanism dynamic confidence fusion, the problem of fusion weight mismatch, high false alarm rate and high miss rate of fault diagnosis caused by time-varying of new energy station environment is solved.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation technology, specifically to a new energy power plant operation and maintenance system and method based on multi-source data fusion. Background Technology

[0002] New energy power plants, including wind farms, photovoltaic power plants, and their complementary power plants, are typically deployed in remote areas. They involve a large number of devices operating in complex environments. To ensure the safe and efficient operation of these plants, existing operation and maintenance systems usually deploy various data acquisition devices, including SCADA systems for collecting equipment operating parameters, CMS systems for collecting vibration status, weather stations for collecting environmental data, and video surveillance and drone inspections for collecting image information. Multi-source data fusion technology, by integrating these heterogeneous data, can more comprehensively assess the health status of equipment and is gradually becoming an important means of intelligent operation and maintenance for new energy power plants.

[0003] However, in existing new energy power plant operation and maintenance solutions based on multi-source data fusion, the fusion weights of each data source are mostly fixed values ​​or set only based on initial prior knowledge. They lack the ability to dynamically adjust according to changes in the actual operating environment and equipment status, resulting in distortion of fusion features when the environment fluctuates or sensors malfunction, leading to false alarms or missed alarms in fault diagnosis and affecting the accuracy of operation and maintenance decisions. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a new energy power station operation and maintenance system and method based on multi-source data fusion, which solves the problems of fusion weight mismatch and high false alarm and false alarm rates in fault diagnosis caused by the time-varying environment of new energy power stations.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: a new energy power station operation and maintenance system based on multi-source data fusion, including a multi-source data acquisition module, an edge data preprocessing module, a feature extraction module, a data fusion engine, a status assessment module, an operation and maintenance decision optimization module, a digital twin visualization module, and a closed-loop feedback module;

[0006] The edge data preprocessing module is communicatively connected to the multi-source data acquisition module. The input of the feature extraction module is connected to the output of the edge data preprocessing module. The input of the data fusion engine is connected to the output of the feature extraction module and a pre-trained initial fusion weight allocation submodule. The state evaluation module is connected to the output of the data fusion engine. The first input of the operation and maintenance decision optimization module is connected to the output of the state evaluation module, and the second input is connected to an operation and maintenance rule base. The digital twin visualization module is communicatively connected to the multi-source data acquisition module, the state evaluation module, and the operation and maintenance decision optimization module. The closed-loop feedback module is connected between the output of the operation and maintenance decision optimization module and the parameter adjustment terminal of the data fusion engine.

[0007] By adopting the above technical solution, a new energy power station operation and maintenance system with a full-link closed-loop self-optimization is obtained. This system solves the data silo problem, improves the accuracy of fault diagnosis and the efficiency of operation and maintenance scheduling, and enables the system to continuously evolve with operational experience. The system uses a multi-source data acquisition module to collect seven types of heterogeneous data, an edge preprocessing module to clean, standardize, and align the data in time, a feature extraction module to generate preliminary fault features, a data fusion engine to dynamically adjust the fusion weights by combining an attention mechanism and closed-loop feedback, a status assessment module to output health level and remaining life prediction, an operation and maintenance decision optimization module to generate multi-objective scheduling schemes, a digital twin visualization module to achieve three-dimensional real-time interaction, and a closed-loop feedback module to correct the fusion parameters by reversing the operation and maintenance execution results.

[0008] Preferably, the multi-source data acquisition module includes: an equipment status acquisition unit, a meteorological station acquisition unit, a video monitoring unit, a drone inspection unit, a power grid communication interface, and a work order management interface, used to collect vibration, temperature, voltage, and current data of the equipment, wind speed, wind direction, and irradiance data of the environment, thermal imaging and visible light video data of the station, image data of equipment appearance defects, power dispatching command data of the grid connection point, and historical operation and maintenance work order data.

[0009] By adopting the above technical solution, and by using equipment status acquisition units, meteorological station acquisition units, video monitoring units, drone inspection units, power grid communication interfaces, and work order management interfaces to collect vibration, temperature, voltage, current, wind speed, wind direction, irradiance, thermal imaging video, visible light video, equipment appearance defect images, grid connection point power dispatch instructions, and historical maintenance work order data, a full-dimensional data source covering the internal status of equipment, external environment, visual appearance, power grid instructions, and historical experience is obtained, providing a complete information foundation for subsequent multimodal fusion.

[0010] Preferably, the data fusion engine includes a dynamic confidence allocation unit based on an attention mechanism. This unit is used to calculate the self-attention weights and mutual attention weights of each modality feature, and adjust the weight coefficients of each data source in feature fusion in real time based on the operation and maintenance effectiveness index received from the closed-loop feedback module. The initial weights of the dynamic confidence allocation unit are provided by the pre-trained initial fusion weight allocation submodule.

[0011] By adopting the above technical solution, since the data fusion engine contains a dynamic confidence allocation unit based on an attention mechanism, this unit calculates the self-attention weight and mutual attention weight of each modality feature, and adjusts the weight coefficients of each data source in feature fusion in real time based on the operation and maintenance effectiveness index received by the closed-loop feedback module. Furthermore, the initial weights are provided by the pre-trained initial fusion weight allocation submodule. Therefore, an adaptive fusion capability is obtained that can both utilize historical prior knowledge and optimize online based on actual operation and maintenance results, enabling the fusion weights to dynamically match the time-varying characteristics of the site.

[0012] Preferably, the state evaluation module employs a hybrid model of deep residual networks and temporal convolutional networks. Its input is the fused feature tensor output by the data fusion engine, and its output includes two parts:

[0013] The first part consists of device health status labels with four levels: normal, attention, abnormal, and severe, and their corresponding probability vectors.

[0014] The second part consists of the predicted remaining useful life of the equipment and the confidence interval.

[0015] By adopting the above technical solution, since the status assessment module uses a hybrid model of deep residual network and temporal convolutional network, its input is a fused feature tensor, and the first part of the output is the equipment health status label and its probability vector containing four levels: normal, attention, abnormal, and severe. The second part is the predicted value of the equipment's remaining service life and the confidence interval. Therefore, it obtains dual assessment results that simultaneously output the current health level and the future service life prediction, providing complete information from the immediate status to the long-term trend for operation and maintenance decisions, and the confidence interval quantifies the prediction uncertainty.

[0016] Preferably, the operation and maintenance decision optimization module takes the lowest operation and maintenance cost and the shortest fault handling time as the objective function, and uses a genetic algorithm to optimize the scheduling scheme of operation and maintenance personnel, vehicles and spare parts. The generated operation and maintenance instructions include: the ID of the device to be maintained, the fault type, the fault severity, the suggested maintenance window period, the list of required spare parts, the standard operating procedure link and the optimal operation and maintenance personnel assignment information.

[0017] By adopting the above technical solution, since the operation and maintenance decision optimization module takes the lowest operation and maintenance cost and the shortest fault handling time as the objective function, it uses a genetic algorithm to optimize the scheduling scheme of operation and maintenance personnel, vehicles and spare parts. The generated operation and maintenance instructions include the ID of the equipment to be maintained, the fault type, the fault severity, the suggested maintenance window period, the list of required spare parts, the standard operating procedure link and the optimal operation and maintenance personnel assignment information. Therefore, the optimal scheduling scheme that balances economy and timeliness is obtained, and the automatic generation of intelligent operation and maintenance work orders under resource constraints is realized.

[0018] The operation and maintenance method for new energy power plants based on multi-source data fusion includes the following steps:

[0019] S1 collects multi-source raw data from new energy power stations through a multi-source data acquisition module;

[0020] S2 performs real-time cleaning, standardization, and spatiotemporal alignment of the data collected by S1 at the edge.

[0021] S3, extract the preliminary fault features corresponding to each data source from the data aligned by S2;

[0022] S4. Based on the features extracted in S3, construct a dynamic confidence fusion model based on the attention mechanism and initialize the fusion weights of each data source.

[0023] S5, based on the dynamic weights calculated in S4, performs weighted feature fusion to generate a fused feature tensor;

[0024] S6, input the fusion feature tensor generated in S5 into the state assessment module, and output the device health status level and the predicted value and confidence interval of the remaining service life.

[0025] S7. Based on the evaluation results of S6 and the configuration of site operation and maintenance resources, generate and execute a multi-objective optimized operation and maintenance scheduling plan.

[0026] S8 collects the operation and maintenance execution results through the closed-loop feedback module and corrects the dynamic confidence fusion model parameters in S4 in reverse.

[0027] By adopting the above technical solution, since the method includes steps S1 to S8: collecting multi-source raw data, cleaning and aligning edge terminals, extracting preliminary fault features, constructing a dynamic confidence fusion model, generating a fusion feature tensor through weighted fusion, outputting health level and remaining lifetime through state assessment, generating and executing a multi-objective optimization scheduling scheme, and correcting fusion model parameters through closed-loop feedback, a complete closed-loop method flow from data acquisition to execution feedback is obtained. The steps form a strong dependency relationship through data flow and parameter updates, ensuring the system's adaptability and long-term stability.

[0028] Preferably, in step S2, the spatiotemporal alignment includes: using GPS timestamps as a reference, using cubic spline interpolation to uniformly resample device status data with a sampling frequency of 1Hz, meteorological data of 10Hz, video data of 25fps and UAV inspection data of 0.1Hz to a 10Hz synchronous data sequence, and establishing a spatial mapping relationship between the data and the digital twin based on the unique device identifier.

[0029] By adopting the above technical solution, since step S2 uses GPS timestamps as a reference and employs cubic spline interpolation to uniformly resample the device status data with a sampling frequency of 1Hz, meteorological data of 10Hz, video data of 25fps, and UAV inspection data of 0.1Hz to a synchronous data sequence of 10Hz, and establishes a spatial mapping relationship between the data and the digital twin based on the unique identifier of the device, strict alignment of multi-source data in time and space is obtained, providing a synchronous and comparable basic dataset for subsequent cross-modal feature fusion.

[0030] Preferably, in step S4, the method for constructing a dynamic confidence fusion model based on an attention mechanism includes: first, calculating the self-attention weights and mutual attention weights of each modality feature to obtain a preliminary attention distribution; then, assigning an initial confidence level provided by a pre-trained model to each data source, and dynamically adjusting the fusion weights of each data source according to the actual health status improvement amount fed back in step S8 using an online update rule, wherein the amount of weight update is positively correlated with the health status improvement amount after the operation and maintenance execution and the current measurement uncertainty of the data source; the final fusion weights are jointly determined by the attention weights and the dynamic confidence level.

[0031] By adopting the above technical solution, since the initial attention distribution is obtained by first calculating the self-attention weight and mutual attention weight of each modality feature in step S4, and then the initial confidence provided by the pre-trained model is assigned to each data source, and the fusion weight is dynamically adjusted according to the actual health status improvement amount fed back in step S8 using online update rules, the update amount of the weight is positively correlated with the health status improvement amount after the operation and maintenance execution and the current measurement uncertainty of the data source. The final fusion weight is jointly determined by the attention weight and the dynamic confidence. Therefore, the dual-driven weight adjustment capability of fusion attention mechanism and reinforcement learning feedback is obtained, which not only pays attention to the instantaneous correlation between modalities, but also accumulates long-term operation and maintenance experience, and introduces measurement uncertainty to achieve anti-noise interference.

[0032] Preferably, in step S5, the weighted feature fusion adopts a combination of adaptive weighted summation and nonlinear activation, which includes: firstly, weighting and summing the dynamic weights given in step S4 with the preliminary fault feature vectors of each data source, and then mapping the summation result through a nonlinear activation function to generate a fused feature tensor.

[0033] By adopting the above technical solution, since step S5 first performs a weighted summation of the dynamic weights and the preliminary fault feature vectors of each data source, and then maps the summation result through a nonlinear activation function to generate a fusion feature tensor, a fusion representation is obtained that retains the interpretability of the weighted linear combination and enhances the feature expression capability through nonlinear activation, thus providing a more discriminative input for the subsequent state assessment module.

[0034] Preferably, in step S8, the reverse correction using the execution result includes: when, within a preset time window after the operation and maintenance work order is executed, if the health status level output in step S6 is detected to have increased by one level or the remaining service life has decreased by more than a preset threshold, the amount of health status improvement is calculated as a positive reward signal to increase the confidence of the high-contribution data source in this fusion; conversely, if there is no improvement or the status deteriorates, a negative reward signal is generated to reduce the confidence of the corresponding data source and trigger the recalibration process of the feature extraction model based on the attention mechanism.

[0035] By adopting the above technical solution, in step S8, when the health status level is improved by one level or the remaining service life decreases by more than a preset threshold within the preset time window after the operation and maintenance work order is executed, the amount of health status improvement is calculated as a positive reward signal to increase the confidence of the high-contribution data source in this fusion; if there is no improvement or the status deteriorates, a negative reward signal is generated to reduce the confidence of the corresponding data source and trigger the recalibration process of the feature extraction model based on the attention mechanism. Therefore, a deep closed-loop optimization capability is obtained that can not only adjust the fusion weights but also actively retrain the feature extraction model, effectively avoiding performance degradation in long-term operation.

[0036] This invention provides a new energy power station operation and maintenance system and method based on multi-source data fusion. It has the following beneficial effects:

[0037] 1. This invention adopts a technical solution that combines real-time cleaning of multi-source heterogeneous data at the edge, spatiotemporal alignment, and dynamic confidence fusion using cloud attention mechanisms. This achieves precise synchronization of multimodal data in time and space, and adaptive fusion with weights that can be adjusted in real time. It achieves the technical effect that the fused feature tensor can accurately reflect the true health status of the equipment and overcome the interference of sensor noise and missing data. Compared with the fault diagnosis schemes in the prior art that use fixed fusion weights or rely on only a single data source, this invention solves the problem of fusion weight mismatch and high false alarm and false negative rates in fault diagnosis caused by the time-varying environment of new energy power stations.

[0038] 2. This invention employs a dual-branch output technology scheme that simultaneously outputs the equipment health status level and its probability vector, as well as the predicted value of remaining service life and its confidence interval, through a status assessment module. This achieves a joint assessment of the current severity of equipment failure and its future degradation trend, providing complete uncertainty quantification information for operation and maintenance decisions. Compared to existing technologies that only output a single health indicator or a point estimate of service life, this invention solves the problem that operation and maintenance personnel cannot predict the remaining service life of equipment and forecast risks, and find it difficult to formulate preventive maintenance plans.

[0039] 3. This invention employs a closed-loop feedback module to reverse-correct the confidence parameters of the data fusion engine based on the operation and maintenance execution results, and triggers the recalibration of the feature extraction model when the feedback fails to improve the results. This achieves a dual self-optimization closed loop for the fusion weights and the feature extraction model, enabling the system to continuously evolve with actual operation and maintenance experience and maintain high accuracy over long-term operation. Compared to the open-loop architecture in existing technologies where the fusion and diagnostic models remain unchanged, this invention solves the problems of lacking self-learning capabilities, being unable to utilize operation and maintenance history to optimize its own performance, and the degradation of model performance after long-term operation. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0041] Figure 2 This is a schematic diagram of the method flow of the present invention;

[0042] Figure 3 This is a schematic diagram of the data preprocessing process of the present invention;

[0043] Figure 4 This is a schematic diagram illustrating the dynamic confidence update of the present invention;

[0044] Figure 5 This is a schematic diagram of the digital twin visualization data flow of the present invention. Detailed Implementation

[0045] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0046] Please see the appendix Figure 1 -Appendix Figure 5 The following is a further description with reference to the embodiments:

[0047] Example 1, as shown in the attached document Figure 1 As shown, this embodiment provides a new energy power station operation and maintenance system based on multi-source data fusion, including the following modules:

[0048] The multi-source data acquisition module includes:

[0049] Equipment Status Acquisition Unit: Connected to the SCADA system of equipment such as photovoltaic inverters, combiner boxes, and transformer substations, it collects data such as voltage, current, power, temperature, and vibration at a sampling frequency of 1Hz.

[0050] Meteorological station data collection unit: Deployed at multiple measuring points within the station, it collects data such as wind speed, wind direction, irradiance, ambient temperature, and humidity at a sampling frequency of 10Hz.

[0051] Video monitoring unit: includes fixed cameras and thermal imagers, which collect thermal imaging video and visible light video of component surfaces at a frame rate of 25fps.

[0052] Drone inspection unit: It automatically performs an inspection once a week, collecting high-definition and infrared images of photovoltaic modules to detect defects such as hot spots, cracks, and dirt, with a sampling frequency of 0.1Hz.

[0053] Power grid communication interface: Obtain power dispatch command data at the grid connection point from the power grid dispatching system via the 104 protocol or Modbus TCP protocol.

[0054] Work order management interface: retrieves historical maintenance work order data from the site operation and maintenance management system, including information such as fault type, processing time, and replacement of spare parts.

[0055] The edge data preprocessing module is deployed on the edge server on the site side and communicates with all the above acquisition units. The edge server uses an Intel Xeon processor and is configured with 64GB of memory and 1TB SSD.

[0056] The preprocessing module comprises three units:

[0057] Data cleaning unit: Outliers are removed using the quartile method. For each data feature, the first quartile Q1 and the third quartile Q3 are calculated. Values ​​less than Q1 - 1.5 × IQR or greater than Q3 + 1.5 × IQR are marked as outliers and removed, where IQR = Q3 - Q1. Missing values ​​are filled using linear interpolation: For the k-th missing point in the time series, the average of the two valid points before and after it is taken as the interpolated value.

[0058] Data standardization unit: Converts data of different dimensions into a unified standard format. For continuous features such as vibration, temperature, and voltage, Z-score standardization is used, with the formula as follows: ;

[0059] Where μ and σ are the mean and standard deviation of the feature in historical data, respectively. For categorical data device models, one-hot encoding is performed.

[0060] Spatiotemporal alignment unit: Based on GPS timestamps, cubic spline interpolation is used to uniformly resample data from different sampling frequencies to 10Hz. Taking 1Hz device status data as an example, a cubic spline function is used. Interpolation is performed on the original discrete points to obtain estimated values ​​at each 10Hz time point. For video frames, inter-frame interpolation is used to generate intermediate frames. For UAV inspection images, the sampling frequency is 0.1Hz. Linear interpolation is used between two inspections to estimate the daily feature change trend, but the images themselves are not interpolated; only the timestamps are aligned. At the same time, based on the device's unique identifier inverter ID and GPS coordinates, each data point is mapped to the corresponding location in the digital twin.

[0061] As attached Figure 3 As shown, the input of the feature extraction module is connected to the output of the edge data preprocessing module. This module extracts preliminary fault features from each type of data:

[0062] Extracting time-series features from SCADA data: including statistics such as mean, variance, peak value, skewness, and kurtosis within a sliding window, with a sliding window length of 60 seconds and a step size of 10 seconds; frequency domain features: extracting the main frequency components and their amplitudes through fast Fourier transform.

[0063] Vibration features are extracted from CMS data, including first harmonic amplitude, second harmonic amplitude, and envelope spectrum features.

[0064] Extracting visual features from video or image data: Extracting 2048-dimensional feature vectors before the fully connected layers using a pre-trained ResNet-50 convolutional neural network.

[0065] Extract environmental features from environmental data, including irradiance fluctuation rate and temperature change rate.

[0066] The data fusion engine receives the preliminary fault feature vectors output by the feature extraction module and incorporates a dynamic confidence allocation unit based on an attention mechanism. This unit first calculates the self-attention weights and mutual attention weights for the input feature matrix. Calculate attention score The attention distribution is obtained, and the initial confidence level of this unit is provided by the pre-trained initial fusion weight allocation submodule during initialization. This pre-training submodule uses a historical fault case dataset and aims to minimize fault diagnosis error. It obtains the initial weights of each modality through offline training using the gradient descent method.

[0067] The closed-loop feedback module connects the output of the operation and maintenance decision optimization module to the parameter adjustment end of the data fusion engine. This module continuously monitors changes in equipment status after operation and maintenance are performed.

[0068] Specifically, in this scheme, the quartile method used in the edge data preprocessing module to remove outliers is based on box plot statistical theory: the interquartile range (IQR) = Q3 - Q1 reflects the dispersion of the middle 50% of the data. The multiplier of 1.5 is a classic threshold proposed by Tukey; under this threshold, the outlier detection rate for normally distributed data is 0.7%, suitable for occasional communication spikes or transient interference in sensor data from new energy power stations. The formula for linear interpolation to fill in missing values ​​is:

[0069] ;

[0070] in and These are the nearest valid values ​​before and after the missing point. The timestamps are for the missing points. This interpolation method has low computational cost and is suitable for real-time processing at edge computing environments. The Z-score normalization formula used in the data normalization unit... In the process, μ and σ are estimated using a rolling update method: the mean and standard deviation within the sliding window are recalculated every 1000 new data points received to adapt to distribution drift caused by equipment aging or seasonal environmental changes. In the feature extraction module, the Fast Fourier Transform formula for frequency domain features is:

[0071] ;

[0072] Where N is the window length, and in this scheme N is 1024. The extracted frequency domain features include 0.5Hz, 1Hz, and second harmonic components. In this scheme, the image feature dimension extracted by ResNet-50 is 2048, which is then reduced to 256 dimensions through a fully connected layer. The parameters of the dimensionality reduction matrix are initialized through transfer learning to reduce the dependence on large-scale labeled images. Through the above principles and parameter selection, the edge preprocessing module can complete high-quality data cleaning and feature extraction with low computational overhead, providing stable and accurate input features for step S3, thereby leading to the subsequent fusion and diagnosis process.

[0073] Example 2, as shown in the attached document Figure 2 As shown, this embodiment provides a new energy power station operation and maintenance method based on multi-source data fusion, including the following steps:

[0074] Perform steps S1 to S3 to collect, preprocess, and extract features in accordance with the system description above.

[0075] Step S4: Construct a dynamic confidence fusion model based on an attention mechanism. Initially, the fusion weights for each data source i are... The pre-trained model provides the following weights: SCADA data weight 0.3, CMS data weight 0.25, video data weight 0.25, and environmental data weight 0.2. Subsequently, the method updates online based on the feedback from step S8, following the following update rules:

[0076] ;

[0077] in, Let be the fusion weight of the i-th data source at time t. The learning rate is set to 0.05 in this embodiment. The improvement in health status after the operation and maintenance is defined as follows: H is the quantitative value of health status, which maps severe to 1, abnormal to 2, attentive to 3, and normal to 4, while the improvement in health status can be positive or negative;

[0078] This represents the measurement uncertainty of the data source at the current moment, derived from the sensor accuracy and historical observation variance, ranging from 0 to 1. The uncertainty for the vibration sensor is 0.3, and the uncertainty for the voltage sensor is 0.05.

[0079] when That is, the weight increases when health status improves; when That is, the weights decrease when the state deteriorates. After the update, normalization is performed to ensure that the sum of all weights is 1. The final fusion weights are jointly determined by the attention weights and the dynamic confidence: the weighted average of the two is taken, where the attention weight accounts for 0.4 and the dynamic confidence accounts for 0.6.

[0080] As attached Figure 4 As shown, step S5: Perform weighted feature fusion. The following formula is used:

[0081] ;

[0082] in This represents the initial fault feature vector of the i-th data source. The dynamic fusion weights given in step S4, It is the hyperbolic tangent activation function.

[0083] Step S6: The state evaluation module uses a hybrid model of deep residual networks and temporal convolutional networks. Input: fused feature tensor After passing through 3 residual blocks and 2 layers of temporal convolutional network, two branches are output:

[0084] Branch 1: Output a 4-dimensional probability vector through a fully connected layer and a Softmax layer, corresponding to four levels: normal, attention, abnormal, and severe.

[0085] Branch 2: Output the remaining lifetime prediction value in hours and the confidence interval, i.e., prediction value ± 1.96 × standard deviation, through a temporal convolutional network with dropout layers.

[0086] Step S7: The operation and maintenance decision optimization module uses the lowest operation and maintenance cost and the shortest fault handling time as the objective function, and employs a genetic algorithm for scheduling optimization. The objective function is:

[0087] ;

[0088] in Total operating costs include personnel hourly costs, vehicle fuel costs, and spare parts costs; This is the sum of the estimated processing times for all faults; and As the weighting coefficient, this embodiment takes... , The genetic algorithm uses chromosome encoding to represent the allocation scheme of maintenance tasks to personnel, vehicles, and spare parts, with a crossover probability of 0.8, a mutation probability of 0.05, a population size of 100, and 100 generations of iteration. The final output maintenance instructions include: the device ID to be maintained, the fault type, the fault severity, the suggested maintenance window, the list of required spare parts, a link to the standard operating procedure, and the optimal maintenance personnel assignment information.

[0089] Step S8: Closed-loop feedback. After the maintenance work order is executed, the system continuously monitors the health status level and remaining service life output in step S6 within a preset time window of 72 hours. If the health status level improves by one level or the remaining service life decreases by more than 20% of a preset threshold, the improvement in health status is calculated. This serves as a positive reward signal, increasing the confidence level of high-contribution data sources in this fusion. Specifically, for each data source, its contribution is calculated. Here, `corr` represents the correlation coefficient between the feature and the health improvement. Data sources with higher contribution levels show a greater increase in confidence. Conversely, if there is no improvement or the condition deteriorates, a negative reward signal is generated, reducing the confidence of the corresponding data source and triggering a recalibration process for the attention-based feature extraction model. The recalibration process includes: extracting normal and faulty data from the historical database for the most recent week, and retraining the query weight matrix in the attention mechanism. Bond weight matrix This allows the model to adapt to changes in the current data distribution.

[0090] Specifically, in this scheme, the dynamic confidence update formula for step S4 is as follows: In, among them, The learning rate controls the sensitivity of the weights to feedback. When the value is 0.05, the maximum single weight change is... It can reach stability after 10 feedbacks, avoiding weight mutations caused by a single abnormal feedback.

[0091] The improvement in health status is obtained by the difference between the quantified health status values ​​before and after maintenance. The quantification mapping uses equally spaced values ​​of 1, 2, 3, and 4. The actual degradation rate of equipment health status is not linear, but to simplify calculations and maintain monotonicity, this scheme uses a linear mapping. Field verification shows that the misjudgment rate is less than 5%.

[0092] The formula for calculating the measurement uncertainty is as follows: ,in, To ensure the sensor's factory precision, the vibration sensor's accuracy is ±5%. The standard deviation of the most recent 100 sampling points. and This formula, representing the maximum value across all data sources, combines instrument error with data variability, which can occur when sensor aging leads to... As the uncertainty increases, the uncertainty automatically rises, thereby reducing the confidence update magnitude of the data source in the fusion process and playing a role in resisting interference.

[0093] The formula for weight normalization is: This step ensures that the weights of all data sources sum to 1, facilitating subsequent weighted summation and the joint weight formula. middle, The preferred value is 0.4, which is determined through cross-validation: testing on historical failure data. From 0.1 to 0.9, a value of 0.4 is chosen to maximize the fault diagnosis F1 score. The calculation principle of the attention weight is based on scaled dot product attention.

[0094] ;

[0095] Where Q, K, and V are the query, key, and value matrices, respectively. denoted as the dimension of the key vector. This scheme employs multi-head attention with 8 heads to capture feature interactions between different subspaces.

[0096] In step S5 The activation function has an output range of It has the advantages of being close to and having a smooth derivative, and its derivative formula is: This effectively mitigates the gradient vanishing problem during backpropagation. Through the aforementioned update rules and fusion methods, the state assessment input obtained in step S6 includes real-time feedback optimization information. Compared to static fusion, the fault diagnosis accuracy is improved by 12%, and the root mean square error of remaining lifetime prediction is reduced by 18%.

[0097] Example 3, as shown in the appendix Figure 5As shown, the digital twin visualization module constructs a 3D digital twin of the new energy power station based on the Unity3D engine. This module communicates with the multi-source data acquisition module, the status assessment module, and the operation and maintenance decision optimization module. It receives the following information in real time:

[0098] The equipment's real-time operating data, including voltage, current, and temperature, is displayed on a 3D model using color mapping: green indicates normal, yellow indicates caution, orange indicates abnormality, and red indicates seriousness.

[0099] Fault location: Highlight the specific location of the faulty device on the geographic information system map;

[0100] Operation and maintenance execution progress: Displays the completion percentage and estimated remaining time of the current operation and maintenance work order in the form of a Gantt chart.

[0101] Maintenance personnel can interact with the digital twin via touchscreen or VR device, and click on any device to view its detailed status, historical data, and predicted lifespan.

[0102] Specifically, in this solution, the data refresh mechanism of the digital twin visualization module adopts a differentiated push strategy: for changes in health status levels, from normal to attention, an immediate high-priority push is triggered with a delay of less than 100ms; for regular updates of operational data, batch pushes are used, with packets sent every 100ms to reduce network overhead. The color mapping is selected based on the sensitivity curve of human vision to distinguish green, yellow, orange, and red. The above RGB values ​​correspond to the standard color gamut and still have good visibility on industrial displays under direct sunlight. The confidence interval for remaining service life is expressed as: predicted value ± 1.96 × standard deviation, where 1.96 corresponds to a 95% confidence level. The width of the shadow band is proportional to the standard deviation. This visualization method allows maintenance personnel to intuitively judge the uncertainty of the prediction, and the wide shadow indicates the need to prepare spare parts in advance. The highlighting of the fault location uses a pulsed halo effect, and its transparency changes periodically over time.

[0103] ;

[0104] Among them, frequency The preferred flicker frequency is 2Hz, as this flicker frequency attracts human attention without causing visual fatigue. The formula for calculating the progress of a work order in a Gantt chart is: Progress Percentage = Sum of completed sub-task hours / Total standard hours × 100%, where the sub-task hours are based on the average of historical work orders of the same type, and this average is stored in the operation and maintenance rule base.

[0105] In the interactive function, the historical data trend curve that pops up when you click on a device uses data from the last 30 days, smoothed using a Savitzky-Golay filter with a window length of 7 and a polynomial order of 2 to reduce random noise while preserving the overall trend of degradation. In this solution, the positional deviation between the digital twin and the real site is controlled within 0.1 meters, achieved through RTK-GPS differential correction. Each key device is equipped with a positioning tag at a positioning frequency of 1Hz. Through this visualization design, maintenance personnel can complete a rapid inspection of all equipment in the field within 30 seconds, significantly improving efficiency compared to the traditional 2-hour manual inspection, and increasing the fault location accuracy to over 95%.

[0106] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A new energy power station operation and maintenance system based on multi-source data fusion, characterized in that, It includes a multi-source data acquisition module, an edge data preprocessing module, a feature extraction module, a data fusion engine, a status assessment module, an operation and maintenance decision optimization module, a digital twin visualization module, and a closed-loop feedback module; The edge data preprocessing module is communicatively connected to the multi-source data acquisition module. The input of the feature extraction module is connected to the output of the edge data preprocessing module. The input of the data fusion engine is connected to the output of the feature extraction module and a pre-trained initial fusion weight allocation submodule. The state evaluation module is connected to the output of the data fusion engine. The first input of the operation and maintenance decision optimization module is connected to the output of the state evaluation module, and the second input is connected to an operation and maintenance rule base. The digital twin visualization module is communicatively connected to the multi-source data acquisition module, the state evaluation module, and the operation and maintenance decision optimization module. The closed-loop feedback module is connected between the output of the operation and maintenance decision optimization module and the parameter adjustment terminal of the data fusion engine.

2. The new energy power station operation and maintenance system based on multi-source data fusion according to claim 1, characterized in that, The multi-source data acquisition module includes: an equipment status acquisition unit, a meteorological station acquisition unit, a video monitoring unit, a drone inspection unit, a power grid communication interface, and a work order management interface. It is used to collect vibration, temperature, voltage, and current data of the equipment, wind speed, wind direction, and irradiance data of the environment, thermal imaging and visible light video data of the station, image data of equipment appearance defects, power dispatching command data of the grid connection point, and historical operation and maintenance work order data.

3. The new energy power station operation and maintenance system based on multi-source data fusion according to claim 1, characterized in that, The data fusion engine contains a dynamic confidence allocation unit based on an attention mechanism. This unit is used to calculate the self-attention weights and mutual attention weights of each modality feature, and adjust the weight coefficients of each data source in feature fusion in real time based on the operation and maintenance effectiveness index received from the closed-loop feedback module. The initial weights of the dynamic confidence allocation unit are provided by the pre-trained initial fusion weight allocation submodule.

4. The new energy power station operation and maintenance system based on multi-source data fusion according to claim 1, characterized in that, The state evaluation module employs a hybrid model combining deep residual networks and temporal convolutional networks. Its input is the fused feature tensor output by the data fusion engine, and its output includes two parts: The first part consists of device health status labels with four levels: normal, attention, abnormal, and severe, and their corresponding probability vectors. The second part consists of the predicted remaining useful life of the equipment and the confidence interval.

5. The new energy power station operation and maintenance system based on multi-source data fusion according to claim 1, characterized in that, The operation and maintenance decision optimization module takes the lowest operation and maintenance cost and the shortest fault handling time as its objective function. It uses a genetic algorithm to optimize the scheduling scheme of operation and maintenance personnel, vehicles and spare parts. The generated operation and maintenance instructions include: the ID of the equipment to be maintained, the fault type, the fault severity, the suggested maintenance window, the list of required spare parts, the standard operating procedure link and the optimal operation and maintenance personnel assignment information.

6. A method for the operation and maintenance of new energy power stations based on multi-source data fusion, characterized in that, Includes the following steps: S1 collects multi-source raw data from new energy power stations through a multi-source data acquisition module; S2 performs real-time cleaning, standardization, and spatiotemporal alignment of the data collected by S1 at the edge. S3, extract the preliminary fault features corresponding to each data source from the data aligned by S2; S4. Based on the features extracted in S3, construct a dynamic confidence fusion model based on the attention mechanism and initialize the fusion weights of each data source. S5, based on the dynamic weights calculated in S4, performs weighted feature fusion to generate a fused feature tensor; S6, input the fusion feature tensor generated in S5 into the state assessment module, and output the device health status level and the predicted value and confidence interval of the remaining service life. S7. Based on the evaluation results of S6 and the configuration of site operation and maintenance resources, generate and execute a multi-objective optimized operation and maintenance scheduling plan. S8 collects the operation and maintenance execution results through the closed-loop feedback module and corrects the dynamic confidence fusion model parameters in S4 in reverse.

7. The new energy power station operation and maintenance method based on multi-source data fusion according to claim 6, characterized in that, In S2, spatiotemporal alignment includes: using GPS timestamps as a reference, using cubic spline interpolation to uniformly resample device status data with a sampling frequency of 1Hz, meteorological data of 10Hz, video data of 25fps and UAV inspection data of 0.1Hz to a 10Hz synchronous data sequence, and establishing a spatial mapping relationship between data and digital twins based on the unique device identifier.

8. The new energy power station operation and maintenance method based on multi-source data fusion according to claim 6, characterized in that, In S4, the method for constructing a dynamic confidence fusion model based on an attention mechanism includes: first, calculating the self-attention weights and mutual attention weights of each modality feature to obtain a preliminary attention distribution; then, assigning an initial confidence level provided by a pre-trained model to each data source, and dynamically adjusting the fusion weights of each data source according to the actual health status improvement amount fed back in S8 using an online update rule, wherein the update amount of the weights is positively correlated with the health status improvement amount after the operation and maintenance execution and the current measurement uncertainty of the data source; the final fusion weights are jointly determined by the attention weights and the dynamic confidence level.

9. The new energy power station operation and maintenance method based on multi-source data fusion according to claim 6, characterized in that, In S5, the weighted feature fusion adopts a combination of adaptive weighted summation and nonlinear activation, including: first, the dynamic weights given in S4 are weighted and summed with the preliminary fault feature vectors of each data source, and then the summation result is mapped through a nonlinear activation function to generate a fused feature tensor.

10. The new energy power station operation and maintenance method based on multi-source data fusion according to claim 6, characterized in that, In S8, the reverse correction using the execution result includes: when the health status level output by S6 is improved by one level or the remaining service life is reduced by more than a preset threshold within a preset time window after the operation and maintenance work order is executed, the amount of health status improvement is calculated as a positive reward signal to increase the confidence of the high-contribution data source in this fusion; conversely, if there is no improvement or the status deteriorates, a negative reward signal is generated to reduce the confidence of the corresponding data source and trigger the recalibration process of the feature extraction model based on the attention mechanism.

Images