Photovoltaic data interpolation and dynamic cleaning decision method based on heterogeneous graph convolution network

Abstract

The application discloses a photovoltaic data interpolation and dynamic cleaning decision method based on a heterogeneous graph convolution network. In the data interpolation stage, the autocorrelation function is used to extract the dominant physical cycle of the photovoltaic power generation sequence and embed the original time sequence characteristics; the node dynamic dependence matrix with time delay is calculated, and the photovoltaic node is divided into key nodes and ordinary nodes; the key nodes extract deep spatial topological features through directed graph convolution, and the ordinary nodes are independently nonlinearly mapped through a multilayer perceptron; the double-layer point convolution fuses the two types of features and reconstructs the missing data. In the cleaning decision stage, the physical evolution of dust accumulation and the economic loss accumulation model are established to determine the static reference cleaning period; the elastic time window is constructed based on the period, the historical sunk cost and the future multi-scale power loss expectation are fused to calculate the comprehensive weight decision value, and the optimal cleaning instruction is triggered within the elastic window. The application considers the data reconstruction accuracy and noise suppression, and realizes the dynamic cleaning scheduling of the maximum net income.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation and intelligent operation and maintenance technology, specifically to a photovoltaic array data interpolation and dynamic cleaning scheduling method that combines heterogeneous graph convolutional neural networks and deep time series prediction. Background Technology

[0002] Against the backdrop of accelerated global energy structure transformation and the continuous advancement of dual-carbon goals, photovoltaic (PV) power generation has become an important component of new energy power generation due to its advantages such as cleanliness, renewability, and flexible deployment. With the continuous expansion of PV grid-connected scale, the refined and digital operation and maintenance management of PV power plants is receiving increasing attention, as its quality directly affects the comprehensive economic benefits throughout the power plant's lifecycle. However, in actual operation, due to uncontrollable factors such as harsh outdoor environments, unstable communication networks, and sensor failures, the multi-source time-series data (such as power generation and meteorological parameters) collected by PV operation and maintenance systems often suffers from large-scale and continuous data gaps. High-quality and complete data are the fundamental prerequisites for conducting accurate PV output prediction, equipment health status assessment, and grid optimization scheduling. Therefore, carrying out high-precision spatiotemporal data reconstruction and interpolation of PV arrays is of significant practical importance for improving the digital management level of power plants and ensuring stable system operation.

[0003] In data monitoring and repair, early typical solutions mainly relied on traditional statistical interpolation or shallow machine learning methods. Statistical interpolation techniques (such as linear interpolation, cubic spline interpolation, moving average, or autoregressive models) are primarily based on the assumptions of local smoothness and continuity of time series, estimating and filling in missing data by constructing simple mathematical fitting curves. Shallow machine learning methods (such as K-Nearest Neighbors (KNN), Support Vector Regression (SVR), or Random Forest) focus on treating data at each time point as independent samples, attempting to extract surface statistical patterns from historical observation records, and thus establish a local mapping relationship between input features and power generation. However, these methods typically treat each photovoltaic module as an isolated node, completely severing the spatial physical topological connections between equipment within the power plant. When faced with the loss of continuous blocks across the entire plant due to communication base station disconnection, the interpolation accuracy of shallow methods lacking spatial depth is extremely limited.

[0004] With the widespread application of deep learning, deep sequence models, represented by Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BiLSTM), and Generative Adversarial Networks (GANs), have been introduced into power data restoration tasks. For example, Yu Xiaoxia et al. proposed a method for restoring missing wind turbine data based on a multi-head attention Bidirectional Long Short-Term Memory (BiLSTM), Zhou Qiuyang et al. combined GAN data augmentation with an improved Bi-LSTM for data reconstruction, and Liu et al. proposed a photovoltaic data interpolation method based on LSTM-WGAN. Although these methods effectively improve the fitting accuracy of single-node temporal dimensions, they essentially still focus on mining the historical feature evolution patterns within a single photovoltaic module, treating each node as an independent sequence, and failing to effectively model the physical topological constraints and spatial coupling relationships between multiple modules in a photovoltaic array. When large-area concurrent missing data occurs, the reconstruction error of pure temporal models remains significantly high due to the lack of spatial dimension information compensation.

[0005] To deeply explore the spatial dependencies of data, recent research has further introduced graph neural networks such as Graph Convolutional Neural Networks (GCN) and Graph Attention Networks (GAT) to construct a joint architecture of spatial graph convolution and temporal networks. Fan et al. proposed a photovoltaic missing data imputation method based on spatiotemporal denoising graph autoencoders, Yu Simiao et al. designed a distributed photovoltaic missing data imputation method based on spatiotemporal convolution-attention networks, and Sivaramkrishnan et al. optimized a multi-view graph imputation framework. These schemes typically define each photovoltaic module as a node in a graph network, utilize the physical spatial distance between nodes to construct a static topology, and aggregate neighborhood features through graph convolution operators to fill in missing values. Research shows that methods based on spatiotemporal fusion architectures significantly outperform single-dimensional methods in reconstruction accuracy.

[0006] However, current photovoltaic data interpolation models based on spatiotemporal fusion architecture still face two major challenges: First, interference from static correlation information. Existing models mostly rely on the static physical spatial distance between devices to construct the topology, which is difficult to characterize the nonlinear dynamic evolution differences caused by local shading, terrain undulations, and meteorological heterogeneity in actual industrial sites, and is very prone to introducing incorrect topological correlation information. Second, the spread of anomalous noise caused by homogeneous mechanisms. Mainstream spatiotemporal reconstruction mechanisms usually treat photovoltaic arrays as homogeneous networks and perform feature aggregation at the same level. Outlier anomalous data generated by large-scale failures of local devices will blindly spread to surrounding healthy nodes along the topological graph through graph convolution aggregation operations, destroying the stable characteristics of the healthy nodes themselves, thereby increasing the reconstruction error of the entire site.

[0007] In terms of cleaning decision-making and scheduling, current mainstream solutions remain at the planned maintenance stage. The cleaning scheduling of the vast majority of photovoltaic modules still relies on manual experience, employing either a crude, static fixed cycle or a simple threshold for power generation efficiency decline. While this model is simple to manage, it completely severs the dynamic connection between module dust accumulation and complex external weather conditions. To improve this situation, some studies have attempted to establish a mapping between environmental parameters and conversion efficiency losses using artificial neural network models (ANN) or machine learning algorithms such as multiple linear regression, or to construct a model of dust accumulation evolution on the module surface by introducing dynamic weather parameters such as temperature, humidity, wind speed, and precipitation. However, most of these solutions are still limited to static loss prediction or local physical modeling of dust accumulation, often simplifying the dust accumulation process to an ideal linear decay, or assuming that the theoretical output power of the module in a clean state is a fixed constant, making it difficult to effectively capture the random disturbances caused by complex weather fluctuations. More importantly, existing research generally lacks a comprehensive consideration of future medium- and long-term weather forecasts and dynamic economic losses, failing to achieve rolling optimization to maximize net benefits in complex industrial environments.

[0008] In summary, the existing photovoltaic operation and maintenance technology has the following shortcomings: (1) When faced with large-area concurrent missing, the spatiotemporal fusion interpolation model based on isomorphic static topology is easily affected by static correlation information interference and abnormal noise diffusion, and the reconstruction accuracy is limited; (2) The cleaning decision lacks the perception of meteorological and ash accumulation dynamic decay, the decision is lagging and the economic benefits are low. Summary of the Invention

[0009] I. Technical Issues The technical problems to be solved by this invention include: (1) Existing photovoltaic time series data repair methods ignore physical topology associations or rely solely on static topology to construct homogeneous networks, resulting in large reconstruction errors of continuously missing data and the blind spread of abnormal noise from large-area failed nodes along the topology. (2) Traditional photovoltaic cleaning strategies lack the perception of weather and dynamic decay of dust accumulation, and adopt fixed cycle or post-event response mode, resulting in decision-making lag and low economic benefits.

[0010] II. Technical Solution To address the aforementioned technical problems, this invention provides a photovoltaic data interpolation and dynamic cleaning decision-making method based on heterogeneous graph convolutional networks, comprising the following steps: Step 1: Obtain multi-source time series data containing missing states collected from photovoltaic power plants, use the autocorrelation function (ACF) to extract the dominant physical periodic features of the photovoltaic power generation sequence, and embed this explicit periodic information into the original time series features to obtain a node feature sequence with fused periodic information; Step 2: Treat the different devices in the photovoltaic array as independent graph nodes, calculate the node dynamic dependency matrix with time delay characteristics, and divide the photovoltaic array nodes into a critical node subset and an ordinary node subset based on the information propagation impact score between nodes. Step 3: For the key node subset, use a directed graph convolutional network with time delay to extract deep spatial topological features; for the ordinary node subset, use a multilayer perceptron (MLP) to perform independent nonlinear mapping on the feature channel dimension of each node. Step 4: Design a two-layer point convolution feature fusion module to concatenate the features output by key nodes and ordinary nodes. Under the guidance of the local interaction matrix, through channel dimensionality reduction and dimensionality increase alignment, the module achieves accurate reconstruction of the local spatiotemporal coupling relationship and outputs the complete interpolated photovoltaic array time series data. Step 5: Based on the interpolated complete time series data, construct a model of the physical evolution of dust accumulation in photovoltaic modules and the accumulation of direct economic losses, calculate the cumulative power generation loss since the last cleaning, and when the cumulative electricity loss first reaches the preset comprehensive cleaning cost per cleaning, establish this time span as the optimal static benchmark cleaning cycle under the current environment. Step 6: Construct an elastic adjustment time window based on the static benchmark cleaning cycle, use a pre-trained deep time series prediction network to obtain future multi-scale (short-term and medium-to-long-term) power generation prediction values, and combine historical sunk costs to calculate the comprehensive weight decision value (DV). Step 7: Monitor the Comprehensive Weight Decision Value (DV) curve daily within the elastic time window. When the DV triggers the preset global optimal solution critical point, the system adaptively issues a scheduling instruction to advance or postpone the cleaning.

[0011] Preferably, in step 1, the multi-source time-series data includes the actual power generation, voltage, and current of the photovoltaic inverter, as well as meteorological data such as ambient temperature and humidity, wind speed, and solar irradiance.

[0012] Preferably, in step 1, the autocorrelation function (ACF) is used to calculate the correlation of the sequence itself under different lag orders, and the lag order corresponding to the significant local peak of the correlation coefficient is selected as the dominant physical period length T; a learnable parameter matrix with dimensions matching the physical period length T and the number of feature channels is constructed, and the time step of the input sequence is moduloed to retrieve the corresponding periodic feature vector, and the periodic feature vector is concatenated and fused with the original time series features in the channel dimension.

[0013] Preferably, in step 2, a node dynamic dependency matrix with time delay features is constructed by calculating the feature similarity between node i at the current time and node j τ time ago. The higher the similarity, the greater the influence of node j on the information propagation of node i. For each node, the sum of its bidirectional propagation similarity with all other nodes is calculated to obtain the information propagation influence score of the node. The top K nodes with the highest scores are selected in descending order to form the key node set, and the remaining nodes are assigned to the ordinary node set.

[0014] Preferably, in step 3, the dynamic dependency matrix is ​​used to characterize the asymmetric spatial flow topology of key nodes, and a directed graph convolutional network with time delay is used to extract deep spatial topological features; for ordinary nodes, their graph convolutional aggregation path is cut off, and a multilayer perceptron is used to perform independent nonlinear mapping on the feature channel dimension of each node.

[0015] Preferably, in step 4, the local interaction matrix serves as a topological guide to constrain information to flow only between nodes with strong physical relationships. The local interaction matrix is ​​generated by thresholding the dynamic dependency matrix. The first layer of pointwise convolution is used to compress and reduce the channel dimension of the concatenated high-dimensional features, and the second layer of pointwise convolution completes feature decoding and re-upgrades the dimension to the original high-dimensional representation space. At the same time, a residual connection mechanism is introduced to aggregate the features extracted by the multi-scale network across layers to prevent the loss of high-frequency dynamic details in the deep network.

[0016] Preferably, in step 5, the dust accumulation process is simulated using physical formulas, and the current dust accumulation is calculated by combining environmental extremes and decay rates; the difference between theoretical power generation and actual power generation is calculated, and the cumulative direct economic loss caused by dust accumulation from the previous cleaning point to the current time is calculated using an integral model, taking into account the current electricity price; when the accumulated electricity loss first reaches the single comprehensive cleaning cost, this time span is established as the optimal static benchmark cleaning cycle under the current environment; wherein, the single comprehensive cleaning cost includes labor costs, equipment depreciation costs, and cleaning resource consumption costs.

[0017] Preferably, in step 6, a KN multiple (e.g., KN=1.1) of the static baseline cleaning cycle is set as the total duration of the elastic decision time window; the expected values ​​of power generation loss in the short and medium term are obtained using a pre-trained deep time series prediction network; the time evolution ratio ρ = t / T_base is introduced to calculate the adaptive evolution weight coefficient, where t is the time span from the current moment to the last cleaning, and T_base is the static baseline cleaning cycle.

[0018] Preferably, in step 6, the formula for calculating the comprehensive weighted decision value DV is: DV(t) = normalize(C_real(t)) · w_history(t) + normalize(E_short(t)) · w_short(t) + normalize(E_mid(t)) · w_mid(t) + PRPDV; where DV(t) is the comprehensive weight decision value calculated daily within the flexible adjustment time window, C_real(t) is the historical cumulative real electricity cost loss value, E_short(t) is the expected value of future short-term ash accumulation power loss, E_mid(t) is the predicted value of future medium- and long-term power generation loss, normalize(·) is the normalization operation, w_history(t), w_short(t), and w_mid(t) are the weight coefficients of the historical real loss item, the short-term expected item, and the medium- and long-term expected item, respectively, and PRPDV is the correction parameter for the manager's subjective operation and maintenance risk preference; the medium- and long-term loss expectation weight w_mid(t) adopts an exponential... The decay function takes into account that the longer the meteorological forecast step size, the greater the error, and the lower the model's confidence level over time. The historical cumulative true loss weight w_history(t) adopts an S-shaped smooth growth model based on the Sigmoid function. The weight of historical true loss accelerates when it approaches half of the baseline period, thereby forcibly constraining the arrival of the cleaning critical point. The short-term loss expectation weight w_short(t) is dynamically obtained by subtracting the historical cumulative true loss weight and the medium- and long-term loss expectation weight from unit 1. By limiting the value range of the initial medium- and long-term weight settings and the maximum upper bound limit parameter of the historical weight, it is ensured that the short-term loss expectation weight always remains in the effective positive range throughout the entire elastic time window.

[0019] Preferably, in step 7, the DV value is calculated daily within the elastic time window and a continuous sensitivity curve is plotted. When the DV value reaches the critical point that makes the overall economic loss per unit area over the entire life cycle reach the lowest point globally, a system alarm is triggered and the final cleaning date scheduling instruction is automatically generated.

[0020] III. Beneficial Effects The beneficial effects of the technical solution of this invention include: (1) In the data interpolation stage, the present invention utilizes a directed graph convolutional network with time delay and a multilayer perceptron to construct a heterogeneous feature extraction branch, and combines it with a two-layer point convolution to perform feature fusion, so that when communication failure causes large-area data loss, it can capture and reconstruct key physical topological features with high spatiotemporal coupling in the photovoltaic array with high precision, while effectively cutting off noise diffusion between irrelevant nodes and improving the robustness of data reconstruction under extreme failure.

[0021] (2) In the cleaning decision-making stage, this invention breaks the traditional fixed-cycle cleaning mode and innovatively integrates the future multi-scale power prediction results with historical sunk costs. Within a flexible time window, it performs rolling optimization through the comprehensive weight decision value (DV). The DV value accurately quantifies the key economic boundary information in the physical evolution process of ash accumulation, enabling the cleaning scheduling decision to adaptively judge and adjust the best cleaning time, thereby maximizing the net benefits of the photovoltaic power plant throughout its entire life cycle.

[0022] (3) The organic combination of the two stages—data interpolation provides high-quality and complete data input for cleaning decisions, and cleaning decisions guide the operation and maintenance rhythm through economic benefit feedback—forms a complete intelligent operation and maintenance closed loop, enhances the operation and maintenance system's ability to capture and understand extreme situations, and perfectly balances the cost of a single cleaning with the nonlinear loss of electricity. Attached Figure Description

[0023] Figure 1 This is the overall system flowchart of the photovoltaic data interpolation and dynamic cleaning decision method based on heterogeneous graph convolutional networks of the present invention. Figure 2 This is a flowchart of the heterogeneous graph convolutional network reconstruction method in this invention; Figure 3 This is the overall flowchart of the cleaning decision-making process in this invention. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0025] I. Data interpolation stage Reference Figure 1 and Figure 2 The specific implementation steps of the data interpolation stage of this invention are as follows: Step 1: Periodic Feature Extraction and Fusion Acquire multi-source time-series data from photovoltaic power plants, including data on missing states. This multi-source time-series data includes the actual power generation, voltage, and current of the photovoltaic inverter, as well as meteorological data such as ambient temperature and humidity, wind speed, and solar irradiance.

[0026] The autocorrelation function (ACF) is used to extract the dominant physical periodic features of photovoltaic power generation sequences, overcoming the limitation of deep networks in stably capturing long-term physical period dependencies. ACF is used to explicitly extract fixed periods to guide the model in learning temporal patterns. The specific implementation is as follows: Step 1.1: Calculate the correlation of the sequence itself at different lag orders using the autocorrelation function (ACF). Select the lag order corresponding to the significant local peak of the correlation coefficient (such as an integer multiple of 24 hours) as the dominant physical period length. The specific formula is as follows:

[0027] Where: time series ,in The total length of the sequence. for The observed value at time t. The sample mean of the sequence is denoted as . .

[0028] Step 1.2: Construct a learnable parameter matrix whose dimensions match the physical cycle length and the number of feature channels, and initialize it as a zero matrix; Step 1.3: During the forward propagation of the network, perform a modulo operation on the time steps of the input sequence and retrieve the corresponding periodic feature vectors from the parameter matrix; Step 1.4: Extract the periodic feature vector Compared with the original time series characteristics The merged features are obtained by splicing and fusing along the channel dimension. : Step 2: Dynamic Dependency Matrix and Node Partitioning Step 2 aims to differentiate the importance of nodes within the power plant, avoiding the introduction of surrounding noise and disruption of the node's stability by forcibly applying graph convolution aggregation when processing ordinary nodes. The specific implementation is as follows: Step 2.1: Construct a dynamic dependency matrix of nodes with time delay characteristics, and calculate the node at the current time. With delay The node before time 1 The feature similarity between nodes; the higher the similarity, the more likely the nodes are to be similar. For nodes The greater the impact of information dissemination, the more specific the formula is as follows:

[0029] Step 2.2: Traverse the dynamic dependency matrix, calculate the sum of the bidirectional propagation similarities of each node with all other nodes, and obtain the information propagation influence score of that node; Step 2.3: Sort all node scores in descending order, select the top K nodes with the highest total scores to form the key node set, and assign the remaining nodes to the ordinary node set: The specific operation is as follows:

[0030]

[0031] Step 3: Heterogeneous splitting feature extraction Step 3 aims to establish a heterogeneous flow-splitting feature extraction mechanism. Traditional homogeneous graph networks indiscriminately aggregate neighborhood features, leading to the blind spread of outlier noise from a large number of failed nodes along the topology. This invention, through flow-splitting processing, preserves the spatiotemporal abrupt change patterns of key nodes while maintaining the independent stability of ordinary nodes. The specific steps are as follows: Step 3.1: For the set of key nodes... Introducing the dynamic dependence matrix with asymmetric flow characteristics calculated in step 2.1. By using directed graph convolution operators with time delays to aggregate spatial neighborhood features, deep dynamic topological features of key nodes under multi-step delays are extracted. The asymmetry of this directed graph convolution is due to the dynamic dependence of the physical directionality of the matrix itself (i.e., the nodes). Past state of nodes i The influence of the current state is naturally reflected. The specific formula for its hierarchical feature transfer is as follows:

[0032] in, For the first Hidden layer features of key nodes in the layer; This is a preset set of time delay orders; In order to delay

[0033] The normalized asymmetric dynamic dependency matrix at time t; For corresponding delay The learnable weight matrix.

[0034] Step 3.2: For ordinary node sets The spatial graph convolutional aggregation path is severed, and its features are directly input into a multilayer perceptron (MLP). Independent nonlinear mappings are performed on the feature channel dimension of each node to obtain representation data that maintains its own stable features. .

[0035] Step 4: Two-layer point convolution feature fusion and data reconstruction Step 4 aims to enable the network to eliminate redundant noise in the low-dimensional manifold space, achieving a deep representation fusion of the dynamic mutation features of key nodes and the stationary features of ordinary nodes. The specific implementation is as follows: Step 4.1: Using the local interaction matrix as a topology guide, constraint information flows only between nodes with strong physical relationships (such as adjacent critical nodes and ordinary nodes). The local interaction matrix is ​​generated by thresholding the dynamic dependency matrix G: Step 4.2: Utilize the first layer of pointwise convolution ( Convolution) on the concatenated high-dimensional features Channel-level compression and dimensionality reduction are performed, and then the input is fed into the feature reconstruction layer. The second layer performs pointwise convolution to complete feature decoding and re-upscale the output features. The process is represented as:

[0036] Step 4.3: Input the dimensionality-reduced features into the feature reconstruction layer, and complete the feature decoding through the second layer of pointwise convolution, re-expanding them to the original high-dimensional representation space. This ensures that ordinary nodes assimilate the dynamic information of adjacent key regions while retaining their own stationary characteristics. Optionally, a residual connection mechanism can be introduced in step 4 to aggregate the features extracted by the multi-scale network across layers to prevent the loss of high-frequency dynamic details in deep networks.

[0037] After steps 1-4, the complete interpolated photovoltaic array timing data is output, completing the first stage of data interpolation.

[0038] II. Cleaning Decision-Making Phase Reference Figure 1 and Figure 3 Based on the complete time-series data output from the first stage, the specific implementation steps of the cleaning decision stage of this invention are as follows: Step 5: Physical Evolution of Ash Accumulation and Establishment of Static Reference Period Based on the complete time-series data output from step 4, a model for the physical evolution of ash accumulation in photovoltaic modules and the cumulative direct economic losses is constructed. A quantitative static evaluation standard is established based on historical operational data to reflect the optimal operation and maintenance rhythm of the power plant under the average ash accumulation rate. The specific implementation is as follows: Step 5.1: Simulate the dust accumulation process using physical formulas, and calculate the current amount of dust accumulation by combining environmental extremes and decay rate; Step 5.2: Calculate the difference between theoretical power generation and actual power generation, and combine it with the current grid electricity price to use an integral model to calculate the cumulative direct economic loss caused by ash blockage from the previous cleaning point to the current time.

[0039] 5.3: Real-time comparison of cumulative economic losses with preset fixed costs for manual cleaning. When the loss amount first exceeds or equals the cleaning cost, output the time interval as the static benchmark cleaning cycle.

[0040] Step 6: Flexible window construction and comprehensive weight decision value calculation Based on the static baseline cleaning cycle established in step 5, a flexible adjustment time window is constructed, integrating future weather and power trends to achieve profit-driven intelligent scheduling. The specific implementation is as follows: The purpose of steps 6 and 7 is to break through the limitations of traditional operation and maintenance, integrate future weather and power trends, and achieve profit-driven intelligent scheduling. The specific implementation is as follows: Step 6.1: Set a specific multiple (e.g., 1.1 times) of the static baseline cleaning cycle as the total duration of the flexible decision-making time window; Step 6.2: Extract the expected values ​​of electricity loss due to ash accumulation in the short and medium to long term; Step 6.3: Introduce the time evolution ratio and calculate the weighting coefficients for adaptive evolution. To quantify the confidence level of data across different time dimensions, weighting coefficients are set based on the current time. t Since the last cleaning time span It exhibits nonlinear dynamic evolution. Let... For the static baseline cleaning cycle established in step 5, the time evolution ratio is defined as follows: The specific evolutionary allocation mechanism is as follows: weight of expected medium- to long-term losses. Considering that the longer the forecast step size, the greater the error, and the lower the model's confidence level over time, an exponential decay function is used to simulate the confidence level decline process.

[0041] in, Set initial medium- to long-term weight values. The preset decay rate hyperparameter. Historical cumulative true loss weights. Since historical losses are sunk costs that have already occurred, their dominant role in cleansing decisions should increase over time. This method employs an S-shaped smooth growth model based on the Sigmoid function.

[0042] in, This is the upper bound parameter for the historical weights. This is the growth slope control parameter. This nonlinear mechanism ensures that the growth slope is controlled near half of the baseline period (i.e., When this happens, the weight of historical actual losses will increase rapidly, thus forcibly constraining the arrival of the cleansing threshold. Short-term expected loss weight. To ensure the physical meaning and normalization characteristics of the comprehensive weights, they are dynamically obtained by subtracting the other two terms from the unit weight of 1.

[0043] During system initialization configuration, by limiting and The range of values ​​is ensured to be within the entire elastic time window. It should always remain within the effective positive range to maintain the ability of short-term high-confidence weather forecasts to respond quickly to decision-making.

[0044] Step 6.4: Introduce the Manager's Subjective Risk Preference Adjustment Parameter (PRPDV) into the formula for flexible fine-tuning of the washout boundary. The specific formula is as follows:

[0045] In the formula, This represents the comprehensive weighted decision value calculated daily within the flexible adjustment time window. This represents the historical cumulative actual electricity cost loss incurred by the system from the last cleaning time to the current calculation time.

[0046] This represents the expected value of dynamic ash accumulation power loss over future short-term timescales obtained using a deep time series prediction network.

[0047] This represents the predicted loss value for future medium- to long-term power generation trends. This refers to the normalization or standardization process performed on various loss or expected data to eliminate scale. These represent the weighting coefficients for the cumulative actual electricity cost loss since the last cleaning, the expected short-term future cost, and the expected medium- and long-term future cost (their values ​​evolve adaptively with the date). This indicates that the reserved adjustment parameters are used to quantify the subjective operational risk preferences of managers for flexible fine-tuning of the cleaning boundary. Finally, in step 7.1, the comprehensive weight decision value curve is monitored daily within the flexible time window. When it is triggered and reaches the preset global minimum loss threshold, the scheduling instruction for early or delayed cleaning is adaptively issued and the cleaning date schedule instruction is automatically generated.

[0048] Step 7: Scrolling optimization and cleaning command issuance The DV value is calculated daily within the flexible time window, and a continuous sensitivity curve is plotted. When the DV value reaches the critical point (optimal decision point) that makes the comprehensive economic loss per unit area over the entire life cycle reach the lowest level globally, the system alarm is triggered and the final cleaning date scheduling instruction is automatically generated. The system adaptively issues scheduling instructions to advance or postpone the cleaning. III. Specific Implementation Examples The following uses a large-scale ground-mounted photovoltaic power station as an example to further illustrate the specific implementation process of the present invention.

[0050] The power station comprises 200 inverters, each connected to multiple photovoltaic strings, corresponding to approximately 1,000 photovoltaic module nodes. During operation, a communication failure resulted in approximately 30% of the node data being continuously lost for over 6 hours. Additionally, the region where the power station is located has an arid climate, leading to a high rate of dust accumulation on the modules, making traditional fixed-cycle cleaning (every 90 days) uneconomical.

[0051] Data interpolation examples Step 1: Collect multi-source time-series data of the power plant over the past 7 days (including power generation, voltage, current, irradiance, temperature and humidity, and wind speed), with approximately 300 nodes having continuous missing segments. Calculate the ACF for each valid power generation sequence, identify the 24-hour period as the dominant physical cycle of photovoltaic power generation, construct a learnable parameter matrix P (dimension 24 × number of feature channels), and concatenate and fuse the periodic feature vector with the original time-series features.

[0052] Step 2: Calculate the dynamic dependency matrix G for all nodes (take the time delay set τ={1,6,12,24} hours), calculate the bidirectional propagation impact score S_i for each node, sort the nodes in descending order and select the Top-50 nodes as the key node set V_key (accounting for 5% of the total number of nodes), and classify the remaining 950 nodes into the normal node set V_normal.

[0053] Step 3: Use a directed graph convolutional network for 50 key nodes, with an asymmetric dynamic dependency matrix as the topology, and perform spatial neighborhood feature aggregation at 4 delay orders; cut off the graph convolutional path for 950 ordinary nodes, and use two layers of MLP to perform independent nonlinear mapping on their respective feature channels.

[0054] Step 4: Concatenate the output features of key nodes and ordinary nodes, and then perform fusion and reconstruction using a double-layer 1×1 convolution (first reducing the dimensionality to 128 and then increasing it back to the original dimensionality). The local interaction matrix M (generated by thresholding G, with a threshold ε=0.3) serves as the topological guide. Residual connections are also introduced to preserve high-frequency dynamic details. The output is 1000 nodes × 7 days of fully imputed time-series data.

[0055] Tests show that the mean absolute error (MAE) of data interpolation in this embodiment is reduced by about 15% compared to the traditional isomorphic GCN method and by about 23% compared to the pure LSTM method.

[0056] Cleaning Decision Implementation Examples Step 5: Based on the interpolated complete data, establish a physical evolution model for dust accumulation. Simulate the cumulative dust amount D(t) using the local average dust decay rate (0.5% / day) and environmental extreme parameters. Calculate the difference between theoretical and actual power generation, and accumulate the economic loss C_real(t) using a rolling integral based on the local grid-connected electricity price of 0.35 yuan / kWh. When C_real(t) first reaches the single comprehensive cleaning cost C_clean = 12000 yuan (including 9000 yuan for labor, 2000 yuan for equipment depreciation, and 1000 yuan for water resources), output the static baseline cleaning cycle T_base = 108 days.

[0057] Step 6: Based on T_base=108 days, set the elastic time window T_window=1.1×108=118.8 days (take 119 days). Use a pre-trained LSTM-Transformer hybrid prediction network to obtain short-term power predictions for the next 7 days and medium-to-long-term power trends for the next 14 days, and calculate the expected short-term loss E_short and expected long-term loss E_mid. Set the initial weight parameters: w_mid0=0.3, α_mid=2.0, L=0.5, k=6.0. Calculate the DV value daily:

[0058] Set the manager's risk preference adjustment parameter PRPDV=0 (neutral preference).

[0059] Step 7: Calculate the DV value daily and plot the curve within the flexible window of days 108-119. On day 116, the DV value reaches the global minimum loss threshold (i.e., the marginal benefit of continuing to wait is lower than the marginal cost), and the system automatically issues a cleaning schedule instruction, determining day 117 as the optimal cleaning date. Compared to the traditional fixed 90-day cycle, this embodiment avoids unnecessary costs caused by premature cleaning.

[0060] The above embodiments of the present invention are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A photovoltaic data interpolation and dynamic cleaning decision-making method based on heterogeneous graph convolutional networks, characterized in that, The process includes the following steps: Step 1: Obtain multi-source time series data containing missing states collected by photovoltaic power plants, use the autocorrelation function (ACF) to extract the dominant physical periodic features of the photovoltaic power generation sequence, and embed this explicit periodic information into the original time series features to obtain a node feature sequence with fused periodic information. Step 2: Treat the different devices in the photovoltaic array as independent graph nodes, calculate the node dynamic dependency matrix with time delay characteristics, and divide the photovoltaic array nodes into a critical node subset and an ordinary node subset based on the information propagation impact score between nodes. Step 3: Extract deep spatial topological features from the subset of key nodes using a directed graph convolutional network with time delay; For the subset of ordinary nodes, a multilayer perceptron (MLP) is used to perform independent nonlinear mapping on the feature channel dimension of each node. Step 4: Design a two-layer point convolution feature fusion module to concatenate the features output by key nodes and ordinary nodes. Under the guidance of the local interaction matrix, through channel dimensionality reduction and dimensionality increase alignment, the module achieves accurate reconstruction of the local spatiotemporal coupling relationship and outputs the complete interpolated photovoltaic array time series data. Step 5: Based on the interpolated complete time series data, construct a model of the physical evolution of dust accumulation in photovoltaic modules and the accumulation of direct economic losses, calculate the cumulative power generation loss since the last cleaning, and when the cumulative electricity loss first reaches the preset comprehensive cleaning cost per cleaning, establish this time span as the optimal static benchmark cleaning cycle under the current environment. Step 6: Construct an elastic adjustment time window based on the optimal static benchmark cleaning cycle, use a pre-trained deep time series prediction network to obtain future multi-scale power generation prediction values, and combine historical sunk costs to calculate the comprehensive weight decision value DV; Step 7: Calculate the DV value daily within the elastic time window and plot a continuous sensitivity curve; When the DV value reaches the critical point that makes the comprehensive economic loss per unit area over the entire life cycle reach the global minimum, trigger a system alarm and automatically generate the final cleaning date scheduling instruction, and the system adaptively issues scheduling instructions to advance or postpone the cleaning.

2. The method according to claim 1, characterized in that, In step 1, the multi-source time-series data includes the actual power generation, voltage, and current of the photovoltaic inverter, as well as meteorological data such as ambient temperature and humidity, wind speed, and solar irradiance.

3. The method according to claim 1, characterized in that, In step 1, the autocorrelation function (ACF) is used to calculate the correlation of the sequence at different lag orders. The lag order corresponding to the significant local peak in the correlation coefficient is selected as the dominant physical period length T. A learnable parameter matrix with dimensions matching the physical period length T and the number of feature channels is constructed. The time steps of the input sequence are moduloed to retrieve the corresponding periodic feature vectors. The periodic feature vectors are then concatenated and fused with the original time-series features along the channel dimension. The specific steps are as follows: Step 1.1: Calculate the correlation of the sequence itself at different lag orders using the autocorrelation function (ACF). Select the lag order corresponding to the significant local peak in the correlation coefficient as the dominant physical period length T. The specific formula is as follows: Where: x is the time series, N is the total length of the series, x_t is the observation value at time t, and the sample mean of the series is denoted as μ; Step 1.2: Construct a learnable parameter matrix whose dimensions match the physical cycle length and the number of feature channels, and initialize it as a zero matrix; Step 1.3: During the forward propagation of the network, perform a modulo operation on the time steps of the input sequence and retrieve the corresponding periodic feature vectors from the parameter matrix; Step 1.4: Extract the periodic feature vector Compared with the original time series characteristics The merged features are obtained by splicing and fusing along the channel dimension. : .

4. The method according to claim 1, characterized in that, In step 2, a node dynamic dependency matrix with time delay is constructed by calculating the feature similarity between node i at the current time and node j τ time moments ago. The higher the similarity, the greater the influence of node j on the information propagation of node i. For each node, the sum of its bidirectional propagation similarity with all other nodes is calculated to obtain the information propagation influence score of that node. The top K nodes with the highest scores are selected in descending order to form the key node set, and the remaining nodes are assigned to the ordinary node set. Different devices in the photovoltaic array are regarded as independent graph nodes. The node dynamic dependency matrix with time delay is calculated, and the photovoltaic array nodes are divided into a key node subset and an ordinary node subset based on the information propagation influence scores between nodes. The specific steps are as follows: Step 2.1: Construct a dynamic dependency matrix of nodes with time delay characteristics, and calculate the node at the current time. With delay The node before time 1 The feature similarity between nodes; the higher the similarity, the more likely the nodes are to be similar. For nodes The greater the impact of information dissemination, the more specific the formula is as follows: Step 2.2: Traverse the dynamic dependency matrix, calculate the sum of the bidirectional propagation similarities of each node with all other nodes, and obtain the information propagation influence score of that node; Step 2.3: Sort all node scores in descending order, select the top-K nodes with the highest total scores to form the key node set, and assign the remaining nodes to the ordinary node set; the specific operation is as follows: ； 。 5. The method according to claim 4, characterized in that, In step 3, the dynamic dependency matrix is ​​used to characterize the asymmetric spatial flow topology of key nodes, and deep spatial topology features are extracted using a directed graph convolutional network with time delay. For ordinary nodes, their graph convolutional aggregation paths are severed, and a multilayer perceptron is used to perform independent nonlinear mapping on the feature channel dimension of each node; the specific steps are as follows: Step 3.1: Targeting the set of key nodes Introducing the dynamic dependence matrix with asymmetric flow characteristics calculated in step 2.

1. By using directed graph convolution operators with time delays to aggregate spatial neighborhood features, deep dynamic topological features of key nodes under multi-step delays are extracted. The asymmetry of this directed graph convolution is dynamically dependent on the physical directionality of the matrix itself, i.e., the nodes. Past state of nodes i The effects of the current state will naturally manifest; That The specific formula for hierarchical feature transfer is as follows: ； in, For the first Hidden layer features of key nodes in the layer; This is a preset set of time delay orders; In order to delay The normalized asymmetric dynamic dependency matrix at time t; For corresponding delay The learnable weight matrix; Step 3.2: For ordinary node sets The spatial graph convolutional aggregation path is severed, and its features are directly input into a multilayer perceptron (MLP). Independent nonlinear mappings are performed on the feature channel dimensions of each node to obtain representation data that maintains its own stable features. .

6. The method according to claim 1, characterized in that, In step 4, the local interaction matrix serves as a topological guide, constraining information to flow only between nodes with strong physical connections. This local interaction matrix is ​​generated by thresholding the dynamic dependency matrix. The first layer of pointwise convolution compresses and reduces the dimensionality of the concatenated high-dimensional features, while the second layer of pointwise convolution decodes the features and restores them to the original high-dimensional representation space. Simultaneously, a residual connection mechanism is introduced to aggregate the features extracted from the multi-scale network across layers. The specific steps are as follows: Step 4.1: Using the local interaction matrix as a topology guide, constraint information flows only between nodes with strong physical connections; the local interaction matrix is ​​generated by thresholding the dynamic dependency matrix G. Step 4.2: Utilize the first layer of pointwise convolution, i.e. Convolution, for concatenated high-dimensional features Channel-level compression and dimensionality reduction are performed, and then the input is fed into the feature reconstruction layer. The second layer performs pointwise convolution to complete feature decoding and re-upscale the output features. The process is represented as: ； Step 4.3: Input the dimensionality-reduced features into the feature reconstruction layer, and complete the feature decoding through the second layer of pointwise convolution. Then, re-upgrade the features to the original high-dimensional representation space, ensuring that ordinary nodes assimilate the dynamic information of adjacent key regions while retaining their own stationary characteristics.

7. The method according to claim 1, characterized in that, In step 5, the dust accumulation process is simulated using physical formulas, and the amount of dust accumulated at the current moment is calculated by combining environmental extremes and decay rate. The difference between theoretical power generation and actual power generation is calculated, and combined with the current electricity price per unit of the power grid, the cumulative direct economic loss caused by ash occlusion from the previous cleaning point to the current time is calculated using an integral model. When the accumulated electricity cost first reaches the comprehensive cost of a single cleaning cycle, this time span is established as the optimal static benchmark cleaning cycle under the current environment; the comprehensive cost of a single cleaning cycle includes labor costs, equipment depreciation costs, and cleaning resource consumption costs; the specific steps are as follows: Step 5.1: Simulate the dust accumulation process using physical formulas, and calculate the current amount of dust accumulation by combining environmental extremes and decay rate; Step 5.2: Calculate the difference between the theoretical power generation and the actual power generation, and combine it with the current electricity price of the power grid, use the integral model to calculate the cumulative direct economic loss caused by ash blockage from the previous cleaning point to the current time. 5.3: Real-time comparison of cumulative economic losses with preset fixed costs for manual cleaning. When the loss amount first exceeds or equals the cleaning cost, output the time interval as the static benchmark cleaning cycle.

8. The method according to claim 7, characterized in that, Step 6 is detailed below: Step 6.1: Set a specific multiple of the static benchmark cleaning cycle as the total duration of the flexible decision time window; Step 6.2: Extract the expected values ​​of electricity loss due to ash accumulation in the short and medium to long term; Step 6.3: Introduce the time evolution ratio and calculate the weighting coefficients for adaptive evolution. To quantify the confidence level of data across different time dimensions, weighting coefficients are set according to the current time. t Since the last cleaning time span It exhibits nonlinear dynamic evolution; assuming For the static baseline cleaning cycle established in step 5, the time evolution ratio is defined as follows: The specific evolutionary allocation mechanism is as follows: weight of expected medium- to long-term losses. The confidence level decrease process is simulated using an exponential decay function: ； in, Set initial medium- to long-term weight values. The preset decay rate hyperparameter; historical cumulative true loss weights. A sigmoid-based smooth growth model is adopted: ； in, This is the upper bound parameter for the historical weights. The growth slope control parameter; short-term loss expectation weight. The result is obtained dynamically by subtracting the other two terms from unit 1: ； During system initialization configuration, by limiting and The range of values ​​is ensured to be within the entire elastic time window. It should always remain within the effective positive range to maintain the ability of short-term high-confidence weather forecasts to respond quickly to decision-making. Step 6.4: Introduce the manager's subjective risk preference adjustment parameter PRPDV into the formula for flexible fine-tuning of the washout boundary. The specific formula is as follows: ; In the formula, This represents the comprehensive weighted decision value calculated daily within the flexible adjustment time window. This represents the historical cumulative actual electricity cost loss incurred by the system from the last cleaning time to the current calculation time. This represents the expected value of dynamic ash accumulation power loss over future short-term timescales obtained using a deep time series prediction network. This represents the predicted loss value for future medium- to long-term power generation trends. This refers to the normalization or standardization process performed on various loss or expected data to eliminate scale. These represent the weighting coefficients for the cumulative actual electricity cost loss since the last cleaning, the expected short-term future cost, and the expected medium- to long-term future cost, respectively. This indicates that the reserved adjustment parameters are used to quantify the subjective operational risk preferences of managers for flexible fine-tuning of the cleaning boundary.

Images