Physical-data dual-driven power prediction method for power generation unit level graph neural network

Abstract

The application relates to the field of power prediction, and discloses a physical-data double-driven power prediction method of a power generation unit level graph neural network, which comprises the following steps: collecting and preprocessing power generation control unit data; grading and fusing meteorological data to generate multi-scale meteorological features; taking a single power generation unit as a node to construct a graph neural network, and defining physical parameters in a graph edge weight as a learnable tensor; generating preliminary power through graph convolution; adaptively fusing the preliminary power and theoretical physical power through a differentiable physical projection layer to obtain corrected power; constructing an augmented Lagrange compliance constraint, taking a power grid standard as a hard constraint; constructing a three-target collaborative loss function containing a prediction error, a physical deviation and a compliance constraint, and calculating and back propagating through a primal-dual alternating strategy gradient. The application realizes deep fusion of physics and data, and significantly improves prediction accuracy, physical consistency and grid compliance.

Description

Technical Field

[0001] This invention relates to the field of power prediction, and more particularly to a power prediction method for power generation units using a graph neural network driven by both physical and data principles. Background Technology

[0002] With the continuous and rapid growth of installed capacity of new energy sources such as wind power and photovoltaics, high-precision power generation prediction has become a key technical means to ensure the safe and stable operation of the power grid and improve the level of new energy consumption. Existing power prediction methods are mainly divided into two categories: physical model methods and data-driven methods. However, the following technical bottlenecks still exist in practical engineering applications: (1) The modeling granularity is coarse, and the spatial correlation between units is not fully captured. Most of the existing mainstream prediction methods take the power station or cluster as the modeling unit and "average" the output characteristics of multiple power generation units, ignoring the differences in output characteristics caused by the aging degree, installation angle, and local environmental differences of different equipment. Another type of method involves single-unit prediction, but usually uses independent time series models (such as LSTM, GRU) to model a single unit independently, without considering the spatial topological correlation between power generation units (such as the wake effect between wind turbine units and the cloud shading propagation effect between photovoltaic strings).

[0003] For example, prior art 1-CN116070525B discloses a method and device for predicting wind farm power generation based on graph attention mechanism. Although it constructs a directed graph with a single turbine as a node and aggregates neighbor features through graph attention, it is purely data-driven, does not introduce physical constraints, and the physical parameters in the graph edge weights are fixed values, which cannot be adaptively calibrated.

[0004] (2) Pure data-driven approaches lack physical constraints and have insufficient extrapolation capabilities under extreme weather conditions. Existing deep learning methods mainly rely on historical data to learn statistical patterns, which are essentially "black box" models. In scenarios with sparse training samples, such as extreme weather, the extrapolation capability decreases significantly. Although some existing technologies introduce physical models, they usually separate physical modeling from data modeling, with physical patterns only used as an independent post-processing verification module.

[0005] For example, prior art 2-CN120996279A discloses a method and system for predicting power efficiency coordination in large-scale wind power bases. It constructs the main power band through an adaptive bandwidth kernel function and combines LSTM to correct the health index. However, it does not use graph neural networks to model the spatial relationship between units, and physical constraints are only used for data cleaning and feature engineering, without being embedded in the network computation graph. The prior art 3-CN121688845A discloses a wind farm power generation prediction and intelligent control system with multi-source data fusion. Although it constructs a digital twin model that includes a physical mechanism layer, the physical model and the data model are still loosely coupled (the physical layer is an independent simulation module), and the physical constraints are not differentiable hard embedded in the neural network training.

[0006] (3) The optimization objectives are out of sync with power grid technical standards, making it difficult to guarantee grid connection compliance. Traditional prediction models often take minimizing the root mean square error (RMSE) or mean absolute error (MAE) as the single optimization objective, ignoring whether the prediction error distribution meets the deviation threshold specified in the power grid connection technical standards. Although some existing technologies have proposed reliability improvement methods based on dynamic loss functions, their optimization objectives are mostly based on the statistical distribution of historical errors, which belongs to probability distribution optimization, rather than compliance constraints targeting standard hard thresholds.

[0007] The prior art 4-CN113591332B discloses a short-term wind power prediction method, device, storage medium and processor, which adopts a weighted average method based on matching meteorological features of similar historical days. It does not involve neural networks, graph structures and physical constraints at all. Its optimization objective is only to minimize the weighted error, which cannot guarantee that the grid connection assessment indicators will be met.

[0008] (4) Fixed physical parameters in graph neural network modeling prevent adaptive calibration. In existing wind power prediction methods based on graph neural networks (such as prior art 1), the edge weights are usually calculated by fixed physical formulas (such as exponential decay functions based on distance and angle). The key physical parameters (such as wake attenuation coefficient and meteorological propagation delay threshold) are set as fixed hyperparameters and cannot be adaptively calibrated according to actual operating data. When there is a deviation between the actual wake characteristics of the wind farm and the theoretical model, the fixed graph structure will introduce systematic errors.

[0009] In summary, there is an urgent need to develop a novel power prediction method that can integrate physical mechanisms, achieve refined modeling at the power generation unit level, meet the hard constraints of power grid technical standards, and has an adaptively calibrated graph structure. Summary of the Invention

[0010] The purpose of this invention is to address the technical shortcomings of existing new energy power prediction technologies, such as coarse modeling granularity leading to insufficient capture of spatial correlation between units, insufficient extrapolation capability of pure data-driven models under extreme weather conditions and lack of physical mechanism constraints, disconnect between optimization objectives and grid connection technology standards leading to non-compliance, and fixed physical parameters in graph neural networks that cannot be adaptively calibrated. This invention provides a physical-data dual-driven graph neural network power prediction method for power generation units.

[0011] Specifically, the present invention provides a power prediction method for a power generation unit-level graph neural network driven by both physical and data aspects, comprising the following steps: S1. Collect multi-dimensional operational data and meteorological data from various power generation control units in the new energy power generation system, and preprocess them to obtain standardized unit characteristic data and meteorological characteristic data. S2. Perform multi-source meteorological hierarchical fusion on the meteorological feature data to generate multi-scale meteorological features; S3. Construct a graph neural network with a single power generation control unit as a node, define the physical parameters in the graph edge weight calculation as learnable physical parameter tensors, and combine the unit feature data with multi-scale meteorological features to construct a dynamic graph neural network. S4. Aggregate neighbor node information through graph convolution operations to generate preliminary predicted power for each unit; S5. Input the preliminary predicted power and the theoretical physical power into the differentiable physical projection layer, and perform adaptive fusion through dynamic confidence gating factor to obtain the corrected predicted power; S6. Construct a compliance constraint target loss based on the augmented Lagrange method, take the grid connection technology standard as a hard constraint, and perform Lagrange multiplier update and penalty coefficient adaptive adjustment. S7. Based on the corrected prediction power, construct a three-objective collaborative loss function that includes prediction error loss, physical projection bias loss and compliance constraint target loss, and perform gradient calculation through the primal-dual alternation strategy. S8. The graph neural network weights, physical projection layer parameters, and the learnable physical parameter tensor are updated synchronously through the backpropagation algorithm, and iterated until convergence to obtain the final predicted power.

[0012] Furthermore, the differentiable physical projection layer in step S5 is adaptively fused using the following formula:

[0013]

[0014] in, The predicted power is corrected by the physical projection layer; The theoretical physical power calculated from the equipment nameplate parameters and real-time meteorological data; The initial predicted power is the output of the graph neural network after graph convolution aggregation; This is a dynamic confidence gating factor. Use the Sigmoid activation function; For graph neural networks at time steps The hidden layer state vector; The weight matrix is ​​a learnable matrix; This is a learnable bias term.

[0015] Furthermore, the construction of the edge weights of the graph neural network in step S3 specifically includes: For wind power scenarios, the weight of the directed edge from upstream node i to downstream node j is calculated using the following formula:

[0016] in, For wind turbine With wind turbine The physical distance between them; This represents the magnitude of the electrical impedance between the grid connection points of the two wind turbines. This refers to the terrain shading coefficient. For the meteorological system from the wind turbine Propagation to wind turbine The time delay is calculated based on wind direction and wind speed; These are distance scale parameters, which can be learned. This is an impedance scale parameter, which can be learned; The time delay threshold is learnable; the learnable parameters constitute a set: ; For photovoltaic scenarios, the calculation is performed using the following formula:

[0017] in, The physical distance between two photovoltaic strings or arrays; The correlation coefficient is the irradiance coefficient. This represents the magnitude of the electrical impedance between the two unit grid connection points; , Similar to wind power scenarios, these are learnable parameters.

[0018] Furthermore, step S6 is as follows: Using the batch-level statistical indicators specified in the grid connection technical standards as hard constraints, a compliance constraint target loss based on the augmented Lagrangian method is constructed:

[0019] in, For Lagrange multipliers; This is the penalty coefficient; For the m-th batch-level statistical indicator, a differentiable approximation is given. For smooth approximation, For approximate accuracy, Use mean absolute error; This is the hard threshold for the m-th indicator as specified in the national standard.

[0020] Furthermore, the differentiable approximation of the batch-level statistical indicators includes: The root mean square error is approximated using a smoothing method, as shown in the following formula:

[0021] in, For the first The prediction error of a sample at a specified time point; Batch size; For sample index within a batch; For the first The prediction error of a sample at a specified time point; To prevent the square root of a negative positive or gradient-exploding positive number; The accuracy is approximately calculated as follows:

[0022] in, For the Sigmoid function; Allowable error bandwidth; The temperature coefficient controls the steepness of the transition. The mean absolute error is as follows: .

[0023] Furthermore, in step S7, the three-objective collaborative loss function Ltotal As shown in the following formula:

[0024] Among them, the first term on the right side of the equation To predict the mean squared error loss, the primary optimization objective is:

[0025] in For installed capacity, The total number of samples; The corrected predicted power; This refers to the actual power. For all learnable weights in a graph neural network; For learnable physical parameters; The second item is the loss due to compliance constraints. LALM It is a hard constraint; the third term Lphy This is the loss due to physical projection deviation, which is a soft constraint. c For adjustment coefficients:

[0026] The optimization process employs a primitive-dual alternating update strategy: the network parameters are updated while the network parameters are updated while the Lagrange multipliers are updated while the network parameters are updated, and the penalty coefficient is adaptively increased when the constraint violation does not decrease.

[0027] Furthermore, the multi-source meteorological hierarchical fusion in step S2 includes: The ultra-short-term forecast adopts a first-level fusion of measured meteorological data at the power generation unit level; where ultra-short-term forecast refers to a forecast period within a first preset range; Short-term forecasts employ a two-stage fusion of numerical weather prediction data; short-term forecasts refer to forecast periods falling within a second preset range. Cluster forecasting employs regional-level spatiotemporal correlation fusion; where cluster forecasting refers to multi-station meteorological fusion forecasting. The fusion results at each level are adaptively weighted and fused using learnable weights. These learnable weights are coupled with graph node features and are dynamically adjusted according to meteorological conditions, as detailed below:

[0028] in, The predicted power is obtained from the three-level fusion. , , For learnable weight parameters, satisfying + + =1.

[0029] A storage medium storing instructions and data for implementing a physical-data dual-driven power generation unit-level graph neural network power prediction method.

[0030] A physical-data dual-driven power generation unit-level graph neural network power prediction device includes: a processor and the storage medium; the processor loads and executes instructions and data in the storage medium to implement a physical-data dual-driven power generation unit-level graph neural network power prediction method.

[0031] The beneficial effects provided by this invention are: (I) Refined Modeling of Dynamic Graph Neural Network at the Power Generation Unit Level This invention constructs a directed graph using a single power generation control unit (wind turbine or photovoltaic string) as the smallest node. The graph edge weights comprehensively consider physical distance, electrical impedance, terrain shading, meteorological propagation delay, and irradiance correlation. Spatial coupling features of upstream neighbors are aggregated through graph convolution operations. Compared with the existing technologies of "site-level averaging" or "single-unit independent time series modeling," this invention explicitly captures the wake effect and cloud shading propagation effect between units, significantly improving prediction accuracy.

[0032] (ii) Differentiable physical projection layer realizes hard embedding of physical laws Unlike existing Physical Information Neural Networks (PINNs), which commonly treat physical constraints as an additional term in the loss function (soft constraints), this invention designs a differentiable physical projection layer as a fixed output layer of the graph neural network. It adaptively fuses data-driven predicted values ​​with theoretical physical power through a dynamic confidence gating factor. The gating factor is dynamically generated by the network's hidden layer states, adaptively switching between data-driven and physical constraint-driven modes under real confidence and uncertain conditions. This design fundamentally solves the problem of extrapolation failure of purely data-driven models under extreme weather conditions, enabling the prediction results to smoothly converge to the physical theoretical boundary.

[0033] (III) The physical parameters of the graph edge weights can be learned to achieve bidirectional closed-loop self-calibration. This invention defines physical parameters such as distance scale parameters, impedance scale parameters, and time delay thresholds in the graph edge weight calculation formula as learnable tensors, initialized with physical prior values, and updated synchronously with network weights via backpropagation during training. This mechanism achieves a two-way closed loop of data-driven correction of physical model parameters and physical constraints guiding network output, enabling the graph topology to dynamically evolve with the actual wake / irradiance propagation characteristics without manual calibration. This solves the problem of fixed physical parameters in existing graph neural networks, which cannot adaptively calibrate.

[0034] (iv) The augmented Lagrange primitive-dual framework realizes the hard constraint satisfaction of grid connection standard. This invention transforms grid connection technical standards into batch-level statistical hard constraints and employs the augmented Lagrange method to construct a primitive-dual optimization framework. To address the issue of non-differentiability of daily accuracy, a Sigmoid differentiable approximation is designed; a minimal constant is added to smooth the RMSE to ensure numerical stability. Through dual gradient ascent of Lagrange multipliers and adaptive scheduling of penalty coefficients, the batch statistics stably meet the national standard threshold upon model convergence, achieving a fundamental breakthrough from soft penalty to hard satisfaction. This solves the engineering pain point of traditional methods, which have small statistical errors but result in grid connection compliance violations.

[0035] (V) Collaborative optimization of three objectives and overall performance improvement This invention constructs a three-objective collaborative loss function comprising prediction error loss (accuracy), physical projection bias loss (physical consistency), and augmented Lagrangian compliance constraint objective (grid connection compliance). Gradient calculation is performed using an alternating primal-dual strategy. The three objectives mutually reinforce each other, forming a positive reinforcement loop: high-confidence predictions from the physical projection layer provide reliable gradient signals for graph parameter self-calibration; the self-calibrated graph structure enhances the feasibility of compliance constraints; and high-quality samples after compliance constraints are met provide feedback to optimize confidence assessment. Experimental results show that, compared to traditional physical methods and pure data-driven GNNs, this invention reduces RMSE (normalized error) by approximately 41%, improves daily accuracy by 7.4 percentage points, and increases the grid connection qualification rate by 14.7 percentage points, achieving synergistic optimization of prediction accuracy, physical consistency, and grid connection compliance.

[0036] (vi) Multi-source meteorological classification fusion and adaptive weighting This invention constructs a three-level meteorological fusion mechanism (measured meteorological data, numerical weather prediction, and regional spatiotemporal correlation), with fusion weights being learnable parameters that are dynamically adjusted according to meteorological conditions and coupled with graph node features. This design enables the use of optimal meteorological sources for ultra-short-term, short-term, and cluster forecasts, further improving the spatiotemporal resolution and robustness of the forecasts. Attached Figure Description

[0037] Figure 1 This is a simplified flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the hardware device of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0039] Before formally describing the present invention, a general description of the solution of the present invention will be given first to facilitate understanding.

[0040] Please refer to Figure 1 The present invention provides a power prediction method for power generation units at the graph neural network level driven by both physical and data aspects, comprising the following steps: S1. Collect multi-dimensional operational data and meteorological data from various power generation control units in the new energy power generation system, and preprocess them to obtain standardized unit characteristic data and meteorological characteristic data. As a specific implementation method, in step S1 of this invention, the power generation control unit refers to the smallest power generation unit in a new energy power generation system that has independent power output capability and spatial location attributes. For wind farms, the power generation control unit is a single wind turbine; for photovoltaic power plants, the power generation control unit is a photovoltaic string or photovoltaic array (an independent power generation unit formed by several photovoltaic modules connected in series, whose output power is connected to the grid through an inverter).

[0041] The multi-dimensional operational data includes static and dynamic characteristic data. Static characteristic data includes: rated power (kW), efficiency parameters (dimensionless), installation tilt angle (degrees), and geographical coordinates (longitude and latitude). Dynamic characteristic data includes: voltage (V), current (A), temperature (°C), and operating status (normal operation, power-limited operation, standby, fault shutdown, etc.). This data is collected in real-time through a SCADA system (Supervisory and Data Acquisition System) at a sampling frequency of 10 seconds to 5 minutes.

[0042] The meteorological data includes: wind speed (m / s), wind direction (degrees), irradiance (W / m²), air temperature (°C), air pressure (hPa), humidity (%), and backsheet temperature (°C). For wind farms, the focus is on collecting wind speed and direction at the hub height; for photovoltaic power plants, the focus is on collecting irradiance on the module surface and backsheet temperature.

[0043] Preprocessing operations include: outlier detection (using the 3σ criterion or box plot method to remove significantly deviating noisy data), missing value imputation (using linear interpolation or forward imputation), and data standardization (using Z-score standardization or Min-Max normalization to map each feature to a uniform numerical range and eliminate dimensional differences). The standardized unit feature data is denoted as a matrix with dimensions (number of units × number of time steps × feature dimension); meteorological feature data also undergoes standardization.

[0044] S2. Perform multi-source meteorological hierarchical fusion on the meteorological feature data to generate multi-scale meteorological features; As a specific implementation method, in step S2 of this invention, a three-level meteorological fusion mechanism is constructed according to the different time scales of the prediction task. Specifically: For ultra-short-term forecasts (forecast period of 0-4 hours), primary fusion is performed using measured meteorological data from the power generation unit. The measured meteorological data originates from sensors installed near the power generation unit, including anemometers, wind vanes, radiometers, and backplane temperature sensors. This level of fusion can capture minute-level meteorological changes, providing high-precision input features for ultra-short-term forecasts.

[0045] For short-term forecasts (forecast period of 0-72 hours), a two-level fusion of numerical weather prediction (NWP) data is used. NWP data originates from meteorological forecast centers and includes gridded forecast data for wind speed, wind direction, temperature, air pressure, and humidity. This level of fusion supports day-ahead planning (such as next-day power generation plan submissions).

[0046] For cluster forecasts (joint forecasts from multiple weather stations), a three-level fusion is performed using regional spatiotemporal correlation. This fusion considers the physical lag effect of meteorological propagation between different weather stations, such as the time delay of cloud movement from station A to station B, or the propagation path of frontal systems within the region.

[0047] The fusion results at each level are adaptively weighted and fused using learnable weights to obtain a multi-scale meteorological feature tensor. This multi-scale meteorological feature serves as the input boundary condition for the subsequent graph neural network model.

[0048] Specifically, the fusion results at each level are adaptively weighted and fused using learnable weights. These learnable weights are coupled with graph node features and are dynamically adjusted according to meteorological conditions, as detailed below:

[0049] in, The predicted power is obtained from the three-level fusion. , , For learnable weight parameters, satisfying + + =1.

[0050] S3. Construct a graph neural network with a single power generation control unit as a node, define the physical parameters in the graph edge weight calculation as learnable physical parameter tensors, and combine the unit feature data with multi-scale meteorological features to construct a dynamic graph neural network. As a specific implementation method, in step S3 of the present invention, a directed graph G = (V, E, A) is constructed, where each node in the node set V represents a power generation control unit, the directed edges in the edge set E represent the physical influence of the upstream unit on the downstream unit (such as the wake effect between wind turbine units, the cloud shading propagation effect between photovoltaic strings), and the element wij of the adjacency matrix A represents the edge weight from node i (upstream) to node j (downstream).

[0051] Unlike traditional graph neural networks that use fixed physical formulas to calculate edge weights, this invention defines the key physical parameters in the edge weight calculation formula as learnable tensors.

[0052] For wind power scenarios, the weight of the directed edge from upstream node i to downstream node j is calculated using the following formula:

[0053] in, For wind turbine With wind turbine The physical distance between them; This represents the magnitude of the electrical impedance between the grid connection points of the two wind turbines. This refers to the terrain shading coefficient. For the meteorological system from the wind turbine Propagation to wind turbine The time delay is calculated based on wind direction and wind speed; These are distance scale parameters, which can be learned. This is an impedance scale parameter, which can be learned; The time delay threshold is learnable; the learnable parameters constitute a set: ; For photovoltaic scenarios, the calculation is performed using the following formula:

[0054] in, The physical distance between two photovoltaic strings or arrays; The correlation coefficient is the irradiance coefficient. This represents the magnitude of the electrical impedance between the two unit grid connection points; , Similar to wind power scenarios, these are learnable parameters.

[0055] Taking wind power as an example, the edge weight calculation formula includes a distance scale parameter. Impedance metric parameters Time delay threshold These parameters are initialized to physical prior values ​​(e.g., ...) at the start of training. Initialized to 8 times the fan rotor diameter, Initialize to the typical impedance value of the collector line. Initialized to the experience propagation delay), and then synchronously updated along with the graph neural network weights via backpropagation during training.

[0056] The specific method for constructing a dynamic graph neural network is as follows: the feature vector of each node i xi It is composed of static features (rated power, efficiency parameters, installation tilt angle, geographical coordinates), dynamic features (voltage, current, temperature, operating status), and multi-scale meteorological features. The graph neural network adopts the architecture of graph convolutional network (GCN) or graph attention network (GAT), with 2-4 layers and 64-256 hidden units per layer.

[0057] For example, a wind farm has two wind turbines, turbine A (upstream) and turbine B (downstream), at a distance of... Electrical impedance Terrain occlusion coefficient (Slight fluctuations), the angle between the wind direction and the line connecting the two machines is 30°, and the wind speed is... Calculated ≈ 43.3 s. Assuming the current... (Current value of learnable parameter) The edge weight, calculated according to the corresponding formula, is 0.053. During training, It will automatically adjust as it propagates backward, making the graph structure more consistent with the actual wake propagation characteristics.

[0058] For example, a photovoltaic power station has two photovoltaic strings, which are located at a distance of... Irradiance correlation coefficient ρij =0.85 (indicating a strong correlation between the two irradiance values), electrical impedance| She | = 0.2 Ω, currently learnable parameter = 120m, = 0.3 Ω, then the edge weight is calculated according to the corresponding formula as: 0.218.

[0059] S4. Aggregate neighbor node information through graph convolution operations to generate preliminary predicted power for each unit; As a specific implementation, in step S4 of the present invention, the graph convolution operation updates the node features in the following manner:

[0060] in, N ( i ) is a node i The set of neighboring nodes (upstream nodes), deg( i ) is a node i The degree; W(l) is the first l The learnable weight matrix of the layer, s It is the ReLU activation function; After multiple layers of graph convolution, the feature vector of each node aggregates the spatial coupling information of its upstream neighbors. Finally, a fully connected layer (MLP) maps the node features to scalar power values, obtaining the preliminary predicted power of each unit. Pgat (Unit: kW). This preliminary power prediction is based solely on data and does not yet incorporate physical constraints.

[0061] S5. Input the preliminary predicted power and the theoretical physical power into the differentiable physical projection layer, and perform adaptive fusion through dynamic confidence gating factor to obtain the corrected predicted power; It should be noted that the differentiable physical projection layer in step S5 undergoes adaptive fusion using the following formula:

[0062]

[0063] in, The predicted power is corrected by the physical projection layer; The theoretical physical power calculated from the equipment nameplate parameters and real-time meteorological data; The initial predicted power is the output of the graph neural network after graph convolution aggregation; This is a dynamic confidence gating factor. Use the Sigmoid activation function; For graph neural networks at time steps The hidden layer state vector; The weight matrix is ​​a learnable matrix; This is a learnable bias term.

[0064] As a specific implementation method, the design of the differentiable physical projection layer in step S5 of this invention is one of the core innovations of this invention. First, the theoretical physical power is calculated based on the equipment nameplate parameters and real-time meteorological data. Pphy : For wind power scenarios: calculations based on the theoretical power curve of wind turbines. Specifically, based on the measured wind speed at the wheel hub height. v Combined with the cut-in wind speed Rated wind speed Cut-off wind speed Rated power air density r Wind turbine radius R Wind energy utilization coefficient The theoretical output power is calculated using parameters such as [parameter 1], and the Sigmoid smoothing function is used to approximate the step boundary to ensure the continuity of gradient backpropagation.

[0065] For photovoltaic scenarios: calculations based on the physical model of photovoltaic modules. Specifically, based on the irradiance received by the component surface. G Nominal conversion efficiency or Effective area A Power temperature coefficient β Cell temperature Standard test conditions reference temperature (25℃), calculate the theoretical output power.

[0066] Then, adaptive fusion is performed according to the following formula:

[0067] in, l The dynamic confidence gating factor, with a value range of [0,1], is generated from the network hidden layer state through a learnable mapping:

[0068] The physical implication of this design is: when the model has a high confidence level regarding the current operating condition (i.e.) (Located in the familiar feature space of the network) l Approaching 0, the model retains its data-driven nonlinear fitting ability; however, when encountering extreme weather or out-of-distribution samples, l Approaching 1, the prediction results smoothly converge to the physical theoretical boundary, thereby improving the extrapolation ability and physical consistency of the model.

[0069] S6. Construct a compliance constraint target loss based on the augmented Lagrange method, take the grid connection technology standard as a hard constraint, and perform Lagrange multiplier update and penalty coefficient adaptive adjustment. It should be noted that step S6 is as follows: Using the batch-level statistical indicators specified in the grid connection technical standards as hard constraints, a compliance constraint target loss based on the augmented Lagrangian method is constructed:

[0070] in, For Lagrange multipliers; This is the penalty coefficient; For the m-th batch-level statistical indicator, a differentiable approximation is given. For smooth approximation, For approximate accuracy, Use mean absolute error; This is the hard threshold for the m-th indicator as specified in the national standard.

[0071] As a specific implementation method, in step S6 of the present invention, the deviation threshold specified in the grid connection technology standard is converted into a hard constraint that the model training must meet.

[0072] Specifically, define the set of constraint indicators: Root mean square error of ultra-short-term forecast in the 4th hour ( The national standard threshold is denoted as e1 ; Short-term forecast accuracy ( The national standard threshold is denoted as e2 (For example, a requirement of no less than 85%) : Absolute deviation of short-term forecasts ( The national standard threshold is denoted as e3 .

[0073] Since the above-mentioned metrics are batch-level statistics (calculated on a small batch) and the daily accuracy is a non-differentiable metric, this invention designs a differentiable approximation function so that these constraint metrics can participate in gradient backpropagation.

[0074] Then, the Lagrange multipliers are introduced. and penalty coefficient Construct the augmented Lagrange function:

[0075] in, For Lagrange multipliers; This is the penalty coefficient; For the first m Differentiable approximation of batch-level statistical indicators For smooth approximation, For approximate accuracy, Use mean absolute error; This is the hard threshold for the m-th indicator as specified in the national standard.

[0076] The differentiable approximation of the batch-level statistical indicators includes: The root mean square error is approximated using a smoothing method, as shown in the following formula:

[0077] in, For the first The prediction error of a sample at a specified time point; Batch size; For sample index within a batch; For the first The prediction error of a sample at a specified time point; To prevent the square root of a negative positive or gradient-exploding positive number; The accuracy is approximately calculated as follows:

[0078] in, For the Sigmoid function; Allowable error bandwidth; The temperature coefficient controls the steepness of the transition. The mean absolute error is as follows: .

[0079] For example: Suppose a batch of data is calculated to obtain... = 0.085 (representing 8.5% of installed capacity), national standard threshold (10% allowed), at this time This item does not incur penalties; like If the sum is 0.12, then max(0, 0.12-0.10) = 0.02, resulting in a secondary penalty. and linear penalty As training progressed, and Adaptive adjustment, forcing Down to the following.

[0080] S7. Based on the corrected prediction power, construct a three-objective collaborative loss function that includes prediction error loss, physical projection bias loss and compliance constraint target loss, and perform gradient calculation through the primal-dual alternation strategy. It should be noted that in step S7, the three-objective collaborative loss function As shown in the following formula:

[0081] As one embodiment, the optimization problem in the theoretical calculations of this invention should be:

[0082] In the formula: This refers to all learnable weights in a graph neural network (including weights of graph convolutional layers, weights of fully connected layers, etc.). For a set of learnable physical parameters (such as ); The corrected predicted power (kW) depends on and ; This represents the actual power (kW).

[0083] To simplify the calculation and practical solution, this invention optimizes the three-objective collaborative loss function. This makes its solution approximate the theoretical value. In other words, the formula... An idealized constrained optimization problem is given, the objective of which is to minimize the prediction error. And it satisfies grid connection standard constraints. To solve this problem in actual training, this invention uses the augmented Lagrangian method to transform the constraints into penalty terms, and additionally introduces a physical deviation regularization term, constructing the following differentiable total loss function:

[0084] Specifically, In the calculation formula, the first term on the right side of the equation To predict the mean squared error loss, the primary optimization objective is:

[0085] in For installed capacity, The total number of samples; The corrected predicted power; This refers to the actual power. For all learnable weights in a graph neural network; For learnable physical parameters; The second item is the loss due to compliance constraints. LALM It is a hard constraint; the third term Lphy This is the loss due to physical projection deviation, which is a soft constraint. c For adjustment coefficients:

[0086] The optimization process employs a primitive-dual alternating update strategy: network parameters are updated while the dual variables are fixed; Lagrange multipliers are updated while the network parameters are fixed; and the penalty coefficient is adaptively increased when a constraint violation does not decrease. During training, the minimum... When the algorithm converges, its solution approximates the theoretical optimal solution.

[0087] As one example, the optimization process employs an alternating primal-dual strategy: in each iteration, the Lagrange multipliers are first fixed. and penalty coefficient Minimize by gradient descent Update the network parameters (original update); then, with the network parameters fixed, perform gradient ascent updates on the Lagrange multipliers (dual update). Repeat this process until... convergence.

[0088] S8. The graph neural network weights, physical projection layer parameters, and the learnable physical parameter tensor are updated synchronously through the backpropagation algorithm, and iterated until convergence to obtain the final predicted power.

[0089] In one specific implementation, step S8 of this invention uses the Adam optimizer (learning rate set to 0.001, decay factor set to 0.9) for gradient backpropagation. The backpropagated gradient flows through the following learnable parameters: Weight matrix of graph neural network (parameters of each graph convolutional layer and fully connected layer). Learnable parameters of the physical projection layer ( and ); The learnable physical parameter tensor in the graph edge weights (for wind power scenarios) For photovoltaic scenarios ).

[0090] During training, each iteration uses a small batch (batch size set to 32-256) of samples to calculate the loss and gradient, updating all the parameters mentioned above. Training stops and the model parameters are saved when the loss function on the validation set no longer decreases for 20 consecutive epochs. The converged model is then used to predict the test set, outputting the final predicted power (unit: kW) for each power generation unit.

[0091] Example 2: This embodiment fully demonstrates the operation process of the solution.

[0092] A certain onshore wind farm is located in North China, with a total installed capacity of 100MW, comprising 50 wind turbine units (numbered #01-#50) with a single unit capacity of 2MW, a hub height of 100m, and a rotor diameter of 120m. The wind farm has an irregular grid layout, spanning 6km east-west and 4km north-south. The terrain is mainly gentle hills, with an elevation variation of 100-300m.

[0093] This embodiment uses the physical-data dual-driven power generation unit-level graph neural network power prediction method proposed in this invention to predict the power generation for the next 0-72 hours, with a time resolution of 15 minutes.

[0094] Step S1: Data Acquisition and Preprocessing Data Acquisition: Historical data for the past 90 days was collected via the wind farm's SCADA system at 15-minute intervals, totaling 8640 time points. The following data was collected for each time point: Power output per fan: (kW), nacelle wind speed v (m / s), cabin wind direction i (°), Generator speed oh (rpm), propeller pitch angle β (°), temperature T (°C), Operating status (0 / 1 / 2 / 3 indicates normal / limited power / standby / fault) Location of each wind turbine: GPS coordinates Accuracy ±1m; Meteorological data: Wind speed, wind direction, temperature, air pressure, and humidity were collected at heights of 10m, 30m, and 100m by two wind measurement towers (located at the northwest and southeast corners of the wind farm); Preprocessing: Outlier removal: The 3σ criterion is used to remove outliers. Outliers comprised approximately 1.2% of the original data. Missing values ​​were filled using linear interpolation.

[0095] Standardization: Perform Z-score standardization on each feature: For example, if the mean wind speed is 7.2 m / s and the standard deviation is 2.5 m / s, then the measured wind speed of 8.5 m / s, after standardization, is (8.5-7.2) / 2.5 = 0.52.

[0096] After processing, the cell feature data matrix is ​​obtained. The time steps are T=96 (corresponding to 24 hours × 4 points / hour), and the feature dimension is d_feat=12.

[0097] Step S2: Multi-source meteorological classification and fusion Three-level fusion computing: Taking the predicted time t=2025-03-15 12:00 (the 4th hour in the future) as an example: First-level fusion (actual measurement): Data from the two wind measurement towers closest to the prediction point are taken and interpolated to the location of each wind turbine using inverse distance weighting. This yields... (Average for the whole game).

[0098] Secondary Numerical Weather Prediction (NWP): This involves obtaining gridded forecast data (3km × 3km resolution) from a regional numerical weather prediction model at that specific moment, interpolating it to the wind farm, and obtaining... .

[0099] Three-level fusion (regional correlation): Based on the historical power time series of upstream wind farm A (30km away), considering the meteorological propagation lag time (approximately 2 hours), we obtain... .

[0100] Learnable weight fusion: The weights learned by the current model are (Updated dynamically according to weather conditions; this is a sunny, low-wind condition). The overall average power corresponding to the fused multi-scale meteorological features is 1853.5 kW. This value is used as the external input feature of the graph neural network and concatenated with the unit features of each node.

[0101] Step S3: Construction of Dynamic Graph Neural Network Graph nodes: 50 wind turbines, feature vector of each node. xi Include: Static characteristics: Rated power 2MW, efficiency coefficient 0.92, installation tilt angle 0° (horizontal axis), geographical coordinates ( x,y,z ); Dynamic characteristics (current moment): nacelle wind speed 8.5 m / s, wind direction 235°, power 1520 kW, speed 1780 rpm, pitch angle 3.2°; Integrated meteorological characteristics (from S2): 1853.5 kW (overall average); Graph edge weight calculation (taking wind power scenario as an example): Select upstream wind turbine #12 (located on the west side) and downstream wind turbine #25 (located on the east side) for calculation: physical distance Electrical impedance Ω (based on collector line parameters); terrain obstruction coefficient (There is a small hill about 50 meters high between the two aircraft); wind direction Azimuth angle of the line connecting the two machines included angle = 150°, cos(150°) = -0.866, since it is a negative value (the wind direction is not from #12 to #25), the actual Treat it as 0, meaning the edge either doesn't exist or has a very small weight. Reselect upstream #07 to downstream #25: The included angle is 235° - 60° = 175°, so cos175° ≈ -0.996, which is still negative. Next, select upstream #18 to downstream #25: The included angle is 235° - 250° = -15°, so cos-15° ≈ 0.966, which is positive. Calculate the time delay. .

[0102] Learnable current values ​​of physical parameters (early training phase): (Initially 8 times the rotor diameter = 960m, but has been reduced through training) ; Substitute into the edge weight formula: ; For all upstream-downstream pairs (where the cosine of the included angle is greater than 0), calculate the edge weights to obtain the adjacency matrix. ; Image convolutional layer: A 2-layer GCN is used, with 128 dimensions in the hidden layer and 1 dimension (power value) in the output layer. The input features are processed through the first convolutional layer to obtain intermediate features. After a second convolution, we get: This is the preliminary power prediction. vector.

[0103] Step S4: Generate preliminary predicted power Perform graph convolution operations to obtain the preliminary predicted power for each wind turbine. For example, for wind turbine #25, its upstream neighbors include seven wind turbines such as #18, #19, and #20. After aggregating their features, .

[0104] Comparison with actual power The initial prediction error is -40 kW (too low). This is because the purely data-driven model does not fully utilize the physical laws and has a bias under conditions where the wake effect is significant.

[0105] Step S5: Differentiable physical projection layer correction Calculate theoretical physical power Pphy : Based on the measured wind speed (8.5 m / s) and power curve of fan #25 (provided by the fan manufacturer), and considering the air density correction (measured air pressure 980 hPa, temperature 12℃, calculated air density ρ = 1.18 kg / m³, standard density 1.225), the theoretical power is obtained from the power curve. .

[0106] Dynamic confidence gating factor : Current hidden layer state (Dimension 128) After linear transformation and Sigmoid activation:

[0107] Assuming the current ,but = 0.231.

[0108] Corrected predicted power: The corrected error was reduced from -40 kW to -24 kW (1696 - 1720 = -24), which is closer to the actual value. Step S6: Augmenting Lagrange Compliance Constraints Constraint calculation (taking the current batch B=64 as an example): (Ultra-short-term 4-hour RMSE): Collect the prediction error for the 4th hour in each batch and calculate the smoothed RMSE: Assumption Therefore, (1 / 64) × 38416 = 600.25, calculated according to the formula. ≈ 24.5kW. National standard threshold. ; Note: In actual standards, RMSE is often calculated as a percentage of installed capacity; this is a simplified example. More reasonable: National standards require... ≤ 10% The current 24.5kW is much smaller than the limit, so the constraint is satisfied.

[0109] (Short-term daily accuracy): Calculation gACC Assuming (15% of a single 2MW unit), temperature coefficient T=0.5. The proportion of prediction errors falling within ±300kW in 64 samples was statistically analyzed, and the result was obtained using the Sigmoid approximation. gACC = 0.92. The national standard requires a daily accuracy rate of ≥ 85%, which is met here (0.92 > 0.85).

[0110] : gMAE = = 35.2 kW, national standard threshold ,satisfy.

[0111] Augmented Lagrange function value: Current Lagrange multipliers Penalty coefficient Since all constraints are satisfied ( ), max(0, ) = 0, therefore: .

[0112] If a constraint is violated, for example Then max(0,500) = 500, which is the quadratic penalty term. This will significantly increase the loss, forcing the model to adjust its parameters to meet the constraints.

[0113] Step S7: Three-objective cooperative loss function Prediction error loss Total installed capacity Current batch sample size N =64, calculate: = 0.0244; Physical projection deviation loss : = 1250 kW², take the adjustment coefficient γ = 0.0001 (to make the magnitude similar to other terms), then · = 0.125.

[0114] Total loss: = 0.0244 + 0.4226 + 0.125 = 0.572.

[0115] Step S8: Backpropagation and Iterative Training The optimizer parameters are set as follows: Adam, learning rate lr=0.001, momentum β1=0.9, β2=0.999.

[0116] Training process: Data from the past 90 days (8640 time points) is used to divide the training, validation, and test sets in an 8:1:1 ratio. The dataset is randomly shuffled in each epoch, with a batch size of 64. After each batch is completed, backpropagation is performed, updating the following parameters: graph convolutional layer weights. Physical projection layer parameters Learnable physical parameters Lagrange multipliers (Dual update) If constraint violation does not improve after 10 consecutive epochs, rm ← 1.5· (upper limit) ).

[0117] After approximately 300 epochs, the loss stabilized at around 0.15, and all constraint indicators met the national standard requirements.

[0118] The overall results are as follows: Test set: March 1 to March 31, 2025 (31 days, 2976 time points, one point every 15 minutes).

[0119] Comparison methods: Method A: Traditional physics method (Jensen wake model + NWP), Method B: Pure data-driven GNN (no physical projection, no constraint optimization), Method C: The method of this invention; The evaluation indicators are: RMSE, MAE, daily accuracy, and grid connection pass rate. The results are shown in Table 1 below: Table 1 Comparison of Results

[0120] This embodiment fully demonstrates the entire process of the method of the present invention, from data acquisition, preprocessing, graph neural network construction, physical projection layer correction, augmented Lagrange hard constraint optimization to final prediction. Experimental results show that, compared with traditional methods and pure data-driven methods, the present invention has significant advantages in prediction accuracy (RMSE reduced by 41%), daily accuracy (improved by 7.4 percentage points), and grid connection assessment pass rate (improved by 14.7 percentage points), fully verifying the effectiveness of the physical-data dual-driven architecture.

[0121] Example 3: Please see Figure 2 , Figure 2 This is a schematic diagram of the hardware device operation according to an embodiment of the present invention. The hardware device specifically includes: a physical-data dual-driven power generation unit-level graph neural network power prediction device 401, a processor 402, and a storage medium 403.

[0122] A physical-data dual-driven power generation unit-level graph neural network power prediction device 401: The physical-data dual-driven power generation unit-level graph neural network power prediction device 401 implements the physical-data dual-driven power generation unit-level graph neural network power prediction method.

[0123] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the physical-data dual-driven power generation unit-level graph neural network power prediction method.

[0124] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the physical-data dual-driven power generation unit-level graph neural network power prediction method.

[0125] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A power prediction method for power generation units using a graph neural network driven by both physical and data aspects, characterized by: Includes the following steps: S1. Collect multi-dimensional operational data and meteorological data from various power generation control units in the new energy power generation system, and preprocess them to obtain standardized unit characteristic data and meteorological characteristic data. S2. Perform multi-source meteorological hierarchical fusion on the meteorological feature data to generate multi-scale meteorological features; S3. Construct a graph neural network with a single power generation control unit as a node, define the physical parameters in the graph edge weight calculation as learnable physical parameter tensors, and combine the unit feature data with multi-scale meteorological features to construct a dynamic graph neural network. S4. Aggregate neighbor node information through graph convolution operations to generate preliminary predicted power for each unit; S5. Input the preliminary predicted power and the theoretical physical power into the differentiable physical projection layer, and perform adaptive fusion through dynamic confidence gating factor to obtain the corrected predicted power; S6. Construct a compliance constraint target loss based on the augmented Lagrange method, take the grid connection technology standard as a hard constraint, and perform Lagrange multiplier update and penalty coefficient adaptive adjustment. S7. Based on the corrected prediction power, construct a three-objective collaborative loss function that includes prediction error loss, physical projection bias loss and compliance constraint target loss, and perform gradient calculation through the primal-dual alternation strategy. S8. The graph neural network weights, physical projection layer parameters, and the learnable physical parameter tensor are updated synchronously through the backpropagation algorithm, and iterated until convergence to obtain the final predicted power.

2. The power prediction method for power generation units at the graph neural network level driven by both physical and data-driven approaches as described in claim 1, characterized in that: The differentiable physical projection layer in step S5 is adaptively fused using the following formula: in, The predicted power is corrected by the physical projection layer; The theoretical physical power calculated from the equipment nameplate parameters and real-time meteorological data; This represents the initial predicted power output of the graph neural network after graph convolution aggregation. It is a dynamic confidence gating factor. Use the Sigmoid activation function; For graph neural networks at time steps The hidden layer state vector; The weight matrix is ​​a learnable matrix; This is a learnable bias term.

3. The power prediction method for power generation unit-level graph neural network driven by both physical and data as described in claim 1, characterized in that: Step S3 involves constructing the edge weights of the graph neural network, specifically including: For wind power scenarios, the weight of the directed edge from upstream node i to downstream node j is calculated using the following formula: in, For wind turbine With wind turbine The physical distance between them; This represents the magnitude of the electrical impedance between the grid connection points of the two wind turbines. This refers to the terrain shading coefficient. For the meteorological system from the wind turbine Propagation to wind turbine The time delay is calculated based on wind direction and wind speed; These are distance scale parameters, which can be learned. This is an impedance scale parameter, which can be learned; The time delay threshold is learnable; the learnable parameters constitute a set: ; For photovoltaic scenarios, the calculation is performed using the following formula: in, The physical distance between two photovoltaic strings or arrays; The correlation coefficient is the irradiance coefficient. This represents the magnitude of the electrical impedance between the two unit grid connection points; , Similar to wind power scenarios, these are learnable parameters.

4. The power prediction method for power generation unit-level graph neural network driven by both physical and data as described in claim 2, characterized in that: Step S6 is as follows: Using the batch-level statistical indicators specified in the grid connection technical standards as hard constraints, a compliance constraint target loss based on the augmented Lagrangian method is constructed: in, For Lagrange multipliers; This is the penalty coefficient; For the m-th batch-level statistical indicator, a differentiable approximation is given. For smooth approximation, For approximate accuracy, Use mean absolute error; This is the hard threshold for the m-th indicator as specified in the national standard.

5. The power prediction method for power generation unit-level graph neural network driven by both physical and data as described in claim 4, characterized in that: The differentiable approximation of the batch-level statistical indicators includes: The root mean square error is approximated using a smoothing method, as shown in the following formula: in, For the first The prediction error of a sample at a specified time point; Batch size; For sample index within a batch; For the first The prediction error of a sample at a specified time point; To prevent the square root of a negative positive or gradient-exploding positive number; The accuracy is approximately calculated as follows: in, For the Sigmoid function; Allowable error bandwidth; The temperature coefficient controls the steepness of the transition. The mean absolute error is as follows: 。 6. The power prediction method for power generation units using a graph neural network driven by both physical and data sources as described in claim 4, characterized in that: In step S7, the three-objective collaborative loss function As shown in the following formula: Among them, the first term on the right side of the equation To predict the mean squared error loss, the primary optimization objective is: in For installed capacity, The total number of samples; The corrected predicted power; This refers to the actual power. For all learnable weights in a graph neural network; For learnable physical parameters; The second item is the loss due to compliance constraints. It is a hard constraint; the third term This is the loss due to physical projection deviation, which is a soft constraint. γ For adjustment coefficients: The optimization process employs a primitive-dual alternating update strategy: the network parameters are updated while the network parameters are updated while the Lagrange multipliers are updated while the network parameters are updated, and the penalty coefficient is adaptively increased when the constraint violation does not decrease.

7. The power prediction method for power generation unit-level graph neural network driven by both physical and data as described in claim 1, characterized in that, The multi-source meteorological classification fusion in step S2 includes: The ultra-short-term forecast adopts a first-level fusion of measured meteorological data at the power generation unit level; where ultra-short-term forecast refers to a forecast period within a first preset range; Short-term forecasts employ a two-stage fusion of numerical weather prediction data; short-term forecasts refer to forecast periods falling within a second preset range. Cluster forecasting employs regional-level spatiotemporal correlation fusion; where cluster forecasting refers to multi-station meteorological fusion forecasting. The fusion results at each level are adaptively weighted and fused using learnable weights. These learnable weights are coupled with graph node features and are dynamically adjusted according to meteorological conditions, as detailed below: in, The predicted power is obtained from the three-level fusion. , , For learnable weight parameters, satisfying + + =1.

8. A storage medium, characterized in that: The storage medium stores instructions and data to implement the power prediction method for a power generation unit-level graph neural network driven by both physical and data as described in any one of claims 1 to 7.

9. A power prediction device for a power generation unit-level graph neural network driven by both physical and data processing, characterized in that: include: A processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement the physical-data dual-driven power generation unit-level graph neural network power prediction method according to any one of claims 1 to 7.

Images