A power distribution network load decomposition prediction method and system based on a physical information neural network

Abstract

The application discloses a power distribution network load decomposition prediction method and system based on a physical information neural network, and relates to the technical field of load prediction. The method comprises the following steps: acquiring measurement data of each monitoring node of a target power distribution network, including total injected power time series data, part of known distributed power output time series data, meteorological time series data and power grid static topology parameters; constructing a physical information neural network model, including decomposing the total injected power into a data-driven subnetwork of natural load and equivalent contribution of distributed power, and a physical constraint subnetwork which is mapped to the physical space of the power grid and calculates a physical loss term; training by using part of the known distributed power data as labels and combining the physical loss term to construct a joint loss function; and inputting real-time measurement data into the trained model to output natural load time series data after stripping the influence of distributed power. The application ensures that the model output is physically interpretable and verifies, and significantly improves the load decomposition accuracy in the data missing scenario.

Description

Technical Field

[0001] This invention relates to the field of load forecasting technology, and in particular to a method and system for load decomposition forecasting of distribution networks based on physical information neural networks. Background Technology

[0002] Distribution network load forecasting is a crucial foundation for power system operation and dispatch. With the large-scale integration of distributed power sources (especially photovoltaic and wind power), the randomness and intermittency of their output lead to severe distortion of measured load data. Traditional forecasting methods based on statistics or simple superposition are unable to accurately isolate the disturbances from new energy sources, posing a serious challenge to the safe and economical operation of the power grid.

[0003] Currently, when processing load forecasting for distribution networks containing distributed generation (DG), the algebraic superposition method is commonly used. This method treats DG output as a "negative load" and removes it from the total injected power through simple addition and subtraction to reconstruct the natural load. However, in complex scenarios with high penetration and multiple access points, DG output data often suffers from problems such as missing measurements, communication delays, or insufficient accuracy. Although there are precedents for applying physical information neural networks to power grid modeling, their physical constraints are limited to electrical laws (such as Kirchhoff's current and voltage laws and generator oscillation equations), failing to consider the physical causal relationship between DG output and meteorological conditions. In the absence of DG data, the inability to effectively utilize meteorological data to reasonably constrain and infer DG output results in a persistent problem of low accuracy in natural load reconstruction. Summary of the Invention

[0004] To address the problem in existing technologies where meteorological data cannot be effectively used to reasonably constrain and infer distributed generation (DG) output when DG data is missing, resulting in low accuracy of natural load reconstruction, this invention provides a distribution network load decomposition and prediction method and system based on a physical information neural network. On one hand, it embeds grid physical laws such as node power balance into the model training; on the other hand, it uses meteorological data to construct physical upper limit constraints on DG output. Even in extreme scenarios with highly scarce DG output data, the model can still infer a reasonable range of DG output through meteorological data, significantly improving the load decomposition accuracy in data-scarce scenarios. The specific technical solution is as follows: This application provides a method for load decomposition and prediction of distribution networks based on physical information neural networks, including the following steps: S1. Acquire measurement data of each monitoring node in the target distribution network; the measurement data includes total injected power time series data, some known distributed power output time series data, meteorological time series data, and grid static topology parameters; the grid static topology parameters include node connection relationships and line impedance parameters; S2. Construct and train a physical information neural network model; the physical information neural network model includes a data-driven subnetwork and a physical constraint subnetwork; The data-driven subnetwork is used to decompose the total injected power time-series data into a first component and a second component; wherein the first component represents the actual natural load and the second component represents the equivalent contribution of distributed power sources. The physical constraint subnetwork includes a first constraint branch and a second constraint branch; the first constraint branch is used to map the output of the data-driven subnetwork to the power grid physical space, and calculate the residual between the output and the power grid physical law to generate a first physical loss term; the second constraint branch is used to calculate the meteorological physical upper limit of the distributed power generation output based on the meteorological time series data, and apply meteorological causal consistency constraints to the second component based on the meteorological physical upper limit to generate a second physical loss term. The physical constraint subnetwork is used to map the output of the data-driven subnetwork to the power grid physical space, and calculate the residual between the output and the power grid physical law to generate a physical loss term; In the training and inference process, the physical information neural network model uses the static topology parameters of the power grid as the fixed input of the first constraint branch of the physical constraint subnetwork. S3. Using the known distributed power output time-series data as partial labels, and combining the first physical loss term and the second physical loss term, a joint loss function is constructed to train the physical information neural network model. S4. Input the real-time measurement data into the trained physical information neural network model, and output the natural load time series data after removing the influence of distributed power sources.

[0005] 2. The distribution network load decomposition prediction method based on physical information neural network according to claim 1, characterized in that the power grid physical laws include node power balance equations and branch power flow constraints; the first physical loss term is specifically the sum of the natural load of each node and the equivalent contribution of distributed generation output by the model, and the mean square error between it and the actual total injected power after power flow calculation using the static topology parameters of the power grid is expressed as: in, This is the first physical loss term; This represents the total number of nodes in the distribution network. For nodes At any moment The actual measured value of the total injected power; Representation based on nodal admittance matrix The power flow calculation function is used to calculate the node. Theoretical injection power; For all nodes output by the model at time... The natural load vector; The output of the model at time 1 The equivalent contribution vector of distributed power sources; The node admittance matrix is ​​constructed from the node connection relationships and line impedance parameters in the static topology parameters of the power grid.

[0006] 3. The distribution network load decomposition prediction method based on physical information neural network according to claim 1, characterized in that the data-driven sub-network adopts an encoder-decoder architecture; wherein the encoder is used to extract deep features of the input time series data, and the decoder includes two parallel output heads, which are used to regress the first component and the second component, respectively.

[0007] 4. The distribution network load decomposition prediction method based on physical information neural network according to claim 1, characterized in that the specific method for generating the second physical loss term in the second constraint branch is as follows: Based on the solar irradiance, ambient temperature, and wind speed in the meteorological time-series data, the theoretical upper limit of the distributed power output of each node is calculated using a distributed power output physical conversion model. The second physical loss item Represented as: in, This represents the total number of nodes in the distribution network. The time window length for time series data; Nodes output by the model At any moment The equivalent contribution value of distributed power sources; For nodes At any moment The theoretical upper limit of output power of distributed power sources; This is a maximum value function used to extract the non-negative part of the value within parentheses.

[0008] 5. The distribution network load decomposition prediction method based on physical information neural network according to claim 1, wherein the joint loss function is expressed as: in, For the joint loss function, This is the data fitting loss term, used to measure the error between the second component of the model output and the partially known distributed power output data. This refers to the first physical loss item; This is the second physical loss term; These are dynamically adjustable weighting coefficients.

[0009] 6. The distribution network load decomposition prediction method based on physical information neural network according to claim 5, characterized in that the dynamically adjustable weighting coefficients... Adaptive adjustments are made based on the missing rate of some known distributed power generation time-series data in the current training data. The adjustment rule is: the higher the missing rate, the higher the weight coefficient of the data fitting loss term. The smaller the value, the higher the weighting coefficient of the first physical loss term. and the weighting coefficient of the second physical loss term The larger the value, the higher the weighting coefficient becomes. When the missing rate exceeds a preset missing rate threshold, the weighting coefficient is updated according to the following rules: in, To preset the basic weight values, To set a preset missing rate threshold, The missing rate, The attenuation coefficient is... This is the growth coefficient.

[0010] Preferably, after inputting real-time measurement data into the trained physical information neural network model and outputting natural load time-series data stripped of the influence of distributed power sources, the method further includes: Based on the aforementioned natural load time series data, dynamic monitoring of abnormal natural load growth is performed; When an anomaly is detected, the corresponding hidden layer features in the physical information neural network model are traced back to locate the main influencing nodes that caused the anomaly and the associated distributed power information.

[0011] Preferably, the data-driven sub-network adopts a parallel structure of temporal convolutional network and multi-head self-attention mechanism: The temporal convolutional network is used to extract local fluctuation features, and the multi-head self-attention mechanism is used to capture long-period dependencies; and the outputs of the temporal convolutional network and the multi-head self-attention mechanism are fused through an adaptive gating unit.

[0012] Preferably, the decoder includes two parallel output heads, respectively used for regressing the first component and the second component, including: The head of the first component is output, and its output is modeled based on the Laplace distribution; wherein the position parameters of the Laplace distribution are obtained by linear transformation of the hidden layer output of the data-driven subnetwork. The header of the second component is output, and its output is modeled based on a Bernoulli-Gaussian mixture distribution; wherein, the mixture distribution includes a Bernoulli binary switching variable and a Gaussian distribution amplitude variable; the Bernoulli binary switching variable is used to characterize the intermittent operation state of the distributed power source, including zero or non-zero output, and the Gaussian distribution amplitude variable is used to regress the specific output amplitude in the non-zero state.

[0013] This application also provides a distribution network load decomposition prediction system based on physical information neural networks, including: The data acquisition module is used to acquire measurement data from each monitoring node of the target distribution network; the measurement data includes total injected power time series data, some known distributed power output time series data, meteorological time series data, and static topology parameters of the power grid. The model building and training module is used to build and train a physical information neural network model. The physical information neural network model includes a data-driven subnetwork and a physical constraint subnetwork. The data-driven subnetwork decomposes the total injected power time-series data into a first component and a second component. The first component represents the actual natural load, and the second component represents the equivalent contribution of distributed generation. The physical constraint subnetwork includes a first constraint branch and a second constraint branch. The first constraint branch maps the output of the data-driven subnetwork to the power grid physical space and calculates the residual between the output and the power grid physical laws, generating a first physical loss term. The second constraint branch calculates the distributed generation based on the meteorological time-series data. The meteorological and physical upper limit of the power source output is determined, and a meteorological causal consistency constraint is applied to the second component based on the meteorological and physical upper limit to generate a second physical loss term. The physical constraint subnetwork is used to map the output of the data-driven subnetwork to the power grid physical space and calculate the residual between the output and the power grid physical law to generate a physical loss term. In the training and inference process, the physical information neural network model uses the static topology parameters of the power grid as the fixed input of the first constraint branch of the physical constraint subnetwork. The partially known distributed power source output time series data is used as partial labels, and combined with the first physical loss term and the second physical loss term, a joint loss function is constructed to train the physical information neural network model. The load forecasting module is used to input real-time measurement data into a trained physical information neural network model and output high-precision natural load time-series data after removing the influence of distributed power sources.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention presents a distribution network load decomposition and prediction method based on a physical information neural network. This method constructs a dual physical constraint subnetwork comprising a first constraint branch and a second constraint branch, achieving a deep integration of electrical physical constraints and meteorological-output causal constraints. The first constraint branch generates electrical physical loss terms based on node power balance equations and branch power flow constraints, ensuring that the model output satisfies the physical laws of the power grid. The second constraint branch calculates the physical upper limit of distributed generation (DG) output based on meteorological time-series data, penalizing any excess output to ensure that DG output conforms to the meteorological-output causal relationship. This joint supervision of dual physical constraints enables the model to implicitly complete data using the power grid topology and physical equations, and to infer reasonable DG output ranges using meteorological data, even in scenarios with missing data. This significantly improves the accuracy of load decomposition and reconstruction, fundamentally overcoming the lack of physical consistency inherent in purely data-driven models. Attached Figure Description

[0015] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0016] Figure 1 A flowchart of a power distribution network load decomposition prediction method based on physical information neural network provided in an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of a power distribution network load decomposition prediction system based on a physical information neural network, provided for an embodiment of the present invention.

[0018] Figure 3 This is a diagram of the physical information neural network model architecture provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] It should be understood that, when used in this specification, the terms “comprising” and “including” indicate the presence of the described feature, integral, step, operation, element and / or component, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0021] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0022] It should also be further understood that the term "and / or" as used in this specification refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes such combinations.

[0023] Please refer to the following examples. Figures 1 to 3 .

[0024] This application provides a distribution network load decomposition prediction method based on a physical information neural network, aiming to solve the problem of low accuracy in natural load reconstruction caused by data gaps and physical inconsistencies in traditional methods under scenarios with a high proportion of distributed power generation. This embodiment's distribution network load decomposition prediction method based on a physical information neural network constructs a neural network model that integrates data-driven and physical constraints. It uses grid topology parameters as fixed inputs for physical constraints, and simultaneously optimizes data fitting errors and physical law residuals during training, ultimately outputting high-precision natural load time-series data. Figure 1 As shown, the method includes the following steps: S1. Acquire measurement data of each monitoring node in the target distribution network; the measurement data includes total injected power time-series data, some known distributed power output time-series data, meteorological time-series data, and grid static topology parameters; the grid static topology parameters include node connection relationships and line impedance parameters; the meteorological time-series data includes solar irradiance, ambient temperature, and wind speed. First, multi-source data were collected from all key monitoring nodes in the target distribution network. These monitoring nodes include feeder heads, low-voltage sides of transformers, distributed generation grid connection points, and important load nodes. The measurement data specifically includes: Total Injection Power Timing Data , representing a node At any moment The net injected active power is usually obtained from the SCADA system, with a sampling interval of 15 minutes or 1 hour; Some known distributed power generation output timing data Accurate measurements are only available for certain nodes or time periods, such as data reported by photovoltaic power plants and monitored residential solar power; data for other nodes or time periods is missing. This data serves as partial labeling for post-training.

[0025] Meteorological time series data This includes factors such as solar irradiance, ambient temperature, and wind speed in the area where the node is located, which are used to construct physical constraints for meteorological causal consistency and to help the model understand the environmental dependence of distributed power output.

[0026] Static topology parameters of the power grid, including node connectivity (adjacency matrix) and line impedance parameters. Transformer turns ratio, etc. These parameters remain constant throughout the process, forming the physical framework of the power grid.

[0027] To improve data quality, the raw data can be preprocessed, including outlier removal, missing value imputation (e.g., using linear interpolation or neighbor mean imputation), and multi-source data alignment. Since the sampling frequencies of SCADA, PMU, and weather forecast data may differ, this application preferably uses neural differential equations to model the continuous time process, mapping all data to the same time grid to ensure time series alignment. Specifically, the neural differential equation uses ordinary neural differential equations (NODE), with the input being the raw sampling sequence of multi-source heterogeneous time series data and the output being uniform time grid data at 15-minute intervals. The NODE contains two hidden layers, each with 64 neurons, using the Tanh activation function, and employs the Euler method to solve the differential equation with a step size of 1 minute.

[0028] In addition, to address the issue of missing data, a distributed power source output feature encoding matrix can be constructed. ,in This indicates that the node always has a reliable DG output tag. This indicates a missing value. This mask will be used in the subsequent loss function to adjust the weights of the data fitting term.

[0029] S2. Construct and train a physical information neural network model; the physical information neural network model includes a data-driven subnetwork and a physical constraint subnetwork; like Figure 3 As shown, the data-driven sub-network is used to decompose the total injected power time-series data into a first component and a second component; wherein, the first component represents the actual natural load, and the second component represents the equivalent contribution of distributed power sources; The physical constraint subnetwork includes a first constraint branch and a second constraint branch; the first constraint branch is used to map the output of the data-driven subnetwork to the power grid physical space, and calculate the residual between the output and the power grid physical law to generate a first physical loss term; the second constraint branch is used to calculate the meteorological physical upper limit of the distributed power generation output based on the meteorological time series data, and apply meteorological causal consistency constraints to the second component based on the meteorological physical upper limit to generate a second physical loss term. In the training and inference process, the physical information neural network model uses the static topology parameters of the power grid as the fixed input of the first constraint branch of the physical constraint subnetwork. By constructing a Physical Information Neural Network (PINN), the model includes two subnetworks: a data-driven subnetwork and a physical constraint subnetwork.

[0030] The data-driven subnetwork is responsible for decomposing the total injected power time series data into two hidden components.

[0031] Specifically, the data-driven sub-network adopts an encoder-decoder architecture, such as... Figure 1 As shown in the text description; wherein, the encoder is used to extract deep features of the input time series data, and the decoder includes two parallel output heads, which are used to regress the first component and the second component, respectively.

[0032] The encoder's input is multivariate time-series data. ,in The time window length, The feature dimensions include total injected power, meteorological data, and timestamps. The encoder consists of a multi-layer temporal convolutional network (TCN), containing three TCN layers with kernel sizes of 3 and dilation rates of 1, 2, and 4, respectively. The hidden layer dimension d=128. Each convolutional layer is followed by batch normalization and a ReLU activation function to progressively extract local temporal features, ultimately outputting the compressed hidden layer representation. .

[0033] The decoder includes two parallel output heads, one for regressing the first component (natural load). ) and the second component (equivalent contribution of distributed power sources) The two output heads share the hidden layer representation from the encoder. However, different fully connected layers are used for mapping.

[0034] The first component output head consists of two fully connected layers with an output dimension of 1 and uses a linear activation function to directly output the predicted value of the natural load. The second component output head also consists of two fully connected layers with an output dimension of 1. Since the output of the distributed power source is non-negative, the output layer uses the ReLU activation function to ensure that the predicted value is non-negative.

[0035] Specifically, the power grid physical laws include node power balance equations and branch power flow constraints; the first physical loss term is specifically the sum of the natural load of each node and the equivalent contribution of distributed generation output by the model, and the mean square error between it and the actual total injected power after power flow calculation using the static topology parameters of the power grid is expressed as: in, This is the first physical loss term; This represents the total number of nodes in the distribution network. For nodes At any moment The actual measured value of the total injected power; Representation based on nodal admittance matrix The power flow calculation function is used to calculate the node. Theoretical injection power; For all nodes output by the model at time... The natural load vector; The output of the model at time 1 The equivalent contribution vector of distributed power sources; The node admittance matrix is ​​constructed from the node connectivity and line impedance parameters in the static topology parameters of the power grid; summation symbol. This represents summing over all nodes.

[0036] The first constraint branch projects the output of the data-driven subnetwork onto the physical constraint space and checks whether it satisfies fundamental physical laws. In this embodiment, the physical constraint subnetwork is based on differentiable flow rate calculation, and its input is: Data-driven subnetwork output of natural loads at each node Equivalent load of DG The sum of the two is the node's net load. ; Power grid topology parameters, node admittance matrix It consists of the node connection relationship and line impedance in step S1, and serves as the fixed input of the physical constraint subnetwork, which does not participate in gradient update.

[0037] The physically constrained subnet performs the following calculations: Based on net load and node admittance matrix The theoretical injected power at each node is obtained by solving the nodal power balance equations. The nodal power balance equations are expressed as follows: in, The node voltage amplitude, For phase angle difference, and These represent the real and imaginary parts of the nodal admittance matrix, respectively. The theoretical injection power of each node is obtained by solving this system of equations using the differentiable Newton-Raphson method. The maximum number of iterations for the differentiable Newton-Raphson method is set to 10. An automatic differentiation framework is used to automatically calculate the Jacobian matrix, and the gradient information of all intermediate variables is retained during the iteration process. When the iteration fails to converge, the damped Newton method is used to correct the step size, and the damping coefficient is set to 0.1. The solution request process is fully differentiable, ensuring that the gradient can propagate back to the entire data-driven subnetwork.

[0038] Comparing theoretical injection power with actual measured values Compare and calculate physical losses : Specifically, the second constraint branch generates the second physical loss term in the following way: Based on the solar irradiance, ambient temperature, and wind speed in the meteorological time-series data, the theoretical upper limit of the distributed power output of each node is calculated using a distributed power output physical conversion model. ; The second physical loss term is expressed as: in, This is the second physical loss term, namely the meteorological factor consistency constraint loss term; This represents the total number of nodes in the distribution network. The time window length for time series data; Nodes output by the model At any moment The equivalent contribution value of distributed power sources; The nodes are calculated based on meteorological data. At any moment The theoretical upper limit of output power of distributed power sources.

[0039] The second constraint branch is based on solar irradiance in meteorological time-series data. Ambient temperature and wind speed The theoretical upper limit of the distributed power output of each node is calculated using a single model of physical conversion of distributed power output. .

[0040] For distributed photovoltaic (PV) power generation, the physical conversion single model is as follows: in, For nodes Photovoltaic conversion efficiency, For the area of ​​the photovoltaic panel, Light irradiance, For temperature coefficient, This is a reference temperature.

[0041] For distributed wind power generation, the physical conversion single model is as follows: in, To cut into wind speed, To cut off the wind speed, Rated wind speed, air density, For the area swept by the wind turbine, The wind energy utilization coefficient, This is the rated power.

[0042] When the DG output of the model exceeds the physical upper limit determined by meteorological conditions, the loss term generates a penalty, guiding the model to learn a physically consistent solution that conforms to the causal relationship between meteorology and output.

[0043] S3. Using the known distributed power output time-series data as partial labels, and combining the first physical loss term and the second physical loss term, a joint loss function is constructed to train the physical information neural network model. This step utilizes some known distributed power output data from step S1 as a supervision signal, and combines it with the physical loss term to construct a joint loss function for end-to-end training of the model.

[0044] In this embodiment, the joint loss function includes data fitting loss, first physical loss, and second physical loss.

[0045] Data fitting loss The error between the model output's DG equivalent contribution and the known labels is measured. Since the labels are only partially present, a masked mean squared error is used: in, The label mask constructed for step S1.

[0046] First physical loss As described in step S2, the second physical loss is calculated from the first constraint branch. Calculated from the second constraint branch. Final joint loss function. Represented as: in, This is the data fitting loss term, used to measure the error between the second component of the model output and the partially known distributed power output data. This refers to the first physical loss item; This is the second physical loss term; These are dynamically adjustable weighting coefficients.

[0047] Specifically, the dynamically adjustable weighting coefficients Adaptive adjustments are made based on the missing rate of some known distributed power generation time-series data in the current training data. The adjustment rule is: the higher the missing rate, the higher the weight coefficient of the data fitting loss term. The smaller the value, the higher the weighting coefficient of the first physical loss term. and the weighting coefficient of the second physical loss term The larger the value, the better.

[0048] In practice, the weighting coefficients are adaptively adjusted based on the missing rate of the current batch number to balance the contributions of data-driven approaches and physical constraints. When the missing rate exceeds a preset missing rate threshold, the weighting coefficients are updated according to the following rules: in, To preset the basic weight values, To set a preset missing rate threshold, The missing rate, The attenuation coefficient is... This is the growth coefficient.

[0049] For example, take , , , , When the missing rate increases, Automatically decrease and The corresponding increase makes the model rely more on physical constraints for learning, avoiding overfitting to noisy labels.

[0050] This embodiment employs a two-stage training strategy to improve convergence stability: In the first stage, the physically constrained subnetwork is frozen, and only the data-driven subnetwork is trained. At this time, the loss function only includes the data fitting loss. (weight) This stage allows the data-driven subnetwork to initially learn the mapping relationship between input and output, making full use of labeled data and training for 20 epochs.

[0051] In the second stage, two sub-networks are trained jointly, introducing a physical loss term and a regularization term. The initial weights are set to... Every 10 epochs, the weights are adjusted based on the physics loss on the validation set, gradually increasing the strength of the physics constraints. Training continues until the validation set loss no longer decreases (stop-wait patience value of 10). Training uses the Adam optimizer, with an initial learning rate of [value missing]. The decay rate is 0.5 times the original value every 20 epochs. The batch size is 32, and the time window length is... (Data from 2 days, with a sampling interval of 1 hour).

[0052] S4. Input the real-time measurement data into the trained physical information neural network model, and output the natural load time series data after removing the influence of distributed power sources.

[0053] Once the model is trained, it can be used for online prediction. Real-time collected measurement data (total injected power, meteorological data, and grid static topology parameters) are input into the model, and forward propagation yields two output components. The first component is the natural load time-series data after removing the influence of distributed generation sources, which can be directly used for subsequent load growth anomaly monitoring, dispatch plan formulation, and other operations.

[0054] It should be noted that the physical constraint subnetwork still participates in the computation during inference, but it no longer calculates the loss; it only provides an implicit guarantee of physical consistency. Due to the existence of the dual physical constraint subnetwork, even if the input data contains noise or some DG data is missing, the model can still output a natural load estimate that conforms to physical laws.

[0055] This invention discloses a distribution network load decomposition and prediction method based on a physical information neural network. By constructing a physical information neural network architecture comprising a data-driven subnetwork and a physical constraint subnetwork, the static topology parameters of the power grid (including node connectivity and line impedance) are used as input to the physical constraint subnetwork during model training and inference, thus integrating data-driven approaches with physical laws. The physical constraint subnetwork includes a first constraint branch and a second constraint branch. The first constraint branch automatically performs calculations of node power balance equations and branch power flow constraints based on the first component of natural load and the second component of equivalent contribution from distributed generation (DG) outputs from the model, generating a physical loss term reflecting the deviation between the output results and the physical laws of the power grid. The second constraint branch calculates the meteorological physical upper limit of DG output based on meteorological time-series data and applies meteorological causal consistency constraints to the second component based on these upper limits, generating a second physical loss term. This fundamentally overcomes the shortcomings of traditional pure data-driven models that disregard physics, ensuring that the residual between the sum of the natural load and the equivalent contribution of DG decomposed by the model and the measured total injected power is minimized after power flow calculation using the actual power grid topology. Simultaneously, the meteorological causal consistency constraints ensure that the DG output output of the model does not exceed the physical upper limit determined by meteorological conditions. Due to the introduction of dual physical constraints, the model not only pursues numerical fitting accuracy during training but is also strongly constrained by the laws of power grid physics and the meteorological-output causality. This ensures that the output natural load time-series data inherently satisfies the node power balance relationship, avoiding physically infeasible solutions caused by data noise or overfitting. Especially in engineering scenarios where distributed generation output data suffers from measurement gaps, communication delays, or insufficient accuracy, the model implicitly completes missing information using power grid topology correlation and physical equations, and infers reasonable ranges for DG output using meteorological data, significantly improving the accuracy of load decomposition and reconstruction.

[0056] Specifically, in a preferred embodiment of this application, after inputting real-time measurement data into the trained physical information neural network model and outputting natural load time-series data stripped of the influence of distributed power sources, the method further includes: Based on the aforementioned natural load time series data, dynamic monitoring of abnormal natural load growth is performed; When an anomaly is detected, the corresponding hidden layer features in the physical information neural network model are traced back to locate the main influencing nodes that caused the anomaly and the associated distributed power information.

[0057] The natural load time series data output in step S4 Real-time monitoring is performed. To adapt to the load characteristics of different nodes and time periods, this embodiment uses an adaptive dynamic threshold method for anomaly detection. For each node... Maintain a length of History sliding window ( (Each hour corresponds to 24 sampling points, with a sampling interval of 1 hour). Calculate the mean of the natural load within the calculation window. and standard deviation : Set threshold coefficient (Take 3, corresponding to the 3σ principle). If the natural load at the current moment satisfies: Then determine the node At any moment An abnormal increase (or decrease) in natural load occurs. The magnitude of the abnormality is recorded. .

[0058] To avoid false alarms caused by load fluctuations, a threshold coefficient is used. It can be dynamically adjusted according to the node type. For example, for industrial load nodes, take... For residential load nodes, take This parameter can be calibrated based on historical data before model deployment.

[0059] When an anomaly is detected, a backtracking mechanism is triggered to analyze the cause of the anomaly. This embodiment uses the integral gradient method to calculate the contribution of input features to the anomaly output.

[0060] For abnormal nodes and abnormal moments The attribution objective is defined as the natural load output by the model. Input features need to be calculated. right The degree of contribution.

[0061] The integral gradient method allocates contributions by accumulating the gradient along the input path from the baseline. The calculation formula is: in: For the first The actual value of each input feature; The baseline value is typically taken as a zero vector or all-zero input (corresponding to reactive power injection and average weather conditions); the integral term is approximated using Riemann summation: exist Uniform sampling in intervals For each point, accumulate the corresponding gradient and take the average.

[0062] The contribution of each input feature is calculated. Then, the following analysis was performed: Sort the input features by absolute value of their contribution and select the top features with the largest contribution. One characteristic.

[0063] By combining feature indexes, the main nodes influencing the anomaly can be located. For example, if the feature with the greatest contribution is a node... At any moment The total injected power indicates the node Historical power variations have a significant impact on the current anomaly.

[0064] Extract input features related to distributed generation (DG) and analyze their contribution. If the contribution of DG-related features is significantly higher, it indicates that the anomaly may be related to distributed generation.

[0065] Based on the attribution analysis results, an anomaly warning report is generated, which includes the location of the abnormal node, the associated distributed power source information, the core rules that triggered the anomaly, and the analysis of the topology impact range.

[0066] By introducing an adaptive dynamic threshold method, adaptive monitoring of load characteristics at different nodes and time periods is achieved, effectively reducing the false alarm rate. When an anomaly is detected, the integral gradient method is used to perform attribution analysis on the input features, transforming the model output results into visualized feature contribution values, and accurately locating the key nodes causing the anomaly and the associated distributed power information.

[0067] Specifically, in a preferred embodiment of this application, the data-driven sub-network adopts a parallel structure of a temporal convolutional network and a multi-head self-attention mechanism: The temporal convolutional network is used to extract local fluctuation features, and the multi-head self-attention mechanism is used to capture long-period dependencies; and the outputs of the temporal convolutional network and the multi-head self-attention mechanism are fused through an adaptive gating unit.

[0068] To further enhance feature extraction capabilities, this embodiment provides an optimized data-driven subnetwork structure, such as... Figure 2 As shown, this structure employs a parallel fusion of temporal convolutional networks and a multi-head self-attention mechanism: Temporal convolutional networks: Input After three layers of dilated causal convolutions, with dilation rates of 1, 2, and 4 in each batch and a kernel size of 3, local fluctuation features of different receptive fields are extracted, and the output is... .

[0069] Multi-head attention branch: Input First, the query is obtained through linear mapping. ,key ,value Then, it captures long-term transient associations through eight self-attention mechanisms and outputs... .

[0070] Adaptive Gated Fusion: The outputs of the two branches are fused through a gating unit. in, For the Sigmoid function, Element-wise multiplication. Features after fusion. It possesses both local fluctuation sensitivity and long-term memory capability, which are injected into the encoder for decomposition and recombination.

[0071] A parallel fusion architecture combining a temporal convolutional network (TCN) and a multi-head self-attention (MSA) mechanism is adopted. The TCN extracts local fluctuation features from different receptive fields through expanded causal convolutions, exhibiting high sensitivity to short-term load fluctuations. The MSA captures long-term dependencies, effectively modeling the diurnal and weekly periodicity of the load. The two are dynamically fused through adaptive gating units, enabling the network to automatically balance the weights of local details and global patterns based on the characteristics of the input data. Compared to a single TCN or Transformer structure, the model architecture in this embodiment significantly improves decomposition accuracy when processing distribution network load data that exhibits both strong fluctuations and strong periodicity.

[0072] Specifically, in a preferred embodiment of this application, the decoder includes two parallel output heads, respectively used for regressing the first component and the second component, including: The head of the first component is output, and its output is modeled based on the Laplace distribution; wherein the position parameters of the Laplace distribution are obtained by linear transformation of the hidden layer output of the data-driven subnetwork. The header of the second component is output, and its output is modeled based on a Bernoulli-Gaussian mixture distribution; wherein, the mixture distribution includes a Bernoulli binary switching variable and a Gaussian distribution amplitude variable; the Bernoulli binary switching variable is used to characterize the intermittent operation state of the distributed power source, including zero or non-zero output, and the Gaussian distribution amplitude variable is used to regress the specific output amplitude in the non-zero state.

[0073] To further enhance the model's ability to characterize the statistical properties of physical quantities, this embodiment provides a preferred output head design, which uses different probability distributions to model natural loads and DG outputs respectively.

[0074] The first component (natural load) output head is modeled based on the Laplace distribution. The network outputs the location parameters of the Laplace distribution. and scale parameters : Position parameters The hidden layer output of the data-driven subnetwork is obtained through a linear transformation, without activation function constraints.

[0075] Scale parameters The output from another linear layer is guaranteed to be positive by the Softplus function.

[0076] The conditional probability density of natural load is: The negative log-likelihood is used as the loss function for this branch: Compared to traditional mean squared error loss, loss based on the Laplace distribution is more robust to outliers, better preserves the true fluctuation details of natural loads, and avoids over-smoothing.

[0077] The second component (DG equivalent load) output head is modeled based on a Bernoulli-Gaussian mixture distribution. The network outputs three sets of parameters: Bernoulli probability , representing the probability that the DG is in an active state (output > 0). The mean of the Gaussian distribution. and variance The output is from another linear layer, where the mean is infinitely bundled and the variance is guaranteed to be positive definite via Softplus. Therefore, the output probability density of the DG is: The loss function for this branch also uses the negative log-likelihood: Combined with the known DG output tag in step S1 (i.e. The system was optimized to accurately match the physical generation process of distributed power sources, significantly improving the prediction accuracy in the zero-value region.

[0078] Specifically, in a preferred embodiment of this application, the joint loss function further includes a rate of change constraint regularization term; the rate of change constraint regularization term is constructed based on the upper limit of the physical ramp rate of the distributed power output and is used to penalize the rate of change of the second component on the time axis.

[0079] like Figure 2 As shown, this application also provides a distribution network load decomposition prediction system based on physical information neural networks, comprising: The data acquisition module is used to acquire measurement data from each monitoring node of the target distribution network; the measurement data includes total injected power time series data, some known distributed power output time series data, meteorological time series data, and static topology parameters of the power grid. The model building and training module is used to build and train a physical information neural network model. The physical information neural network model includes a data-driven subnetwork and a physical constraint subnetwork. The data-driven subnetwork decomposes the total injected power time-series data into a first component and a second component. The first component represents the actual natural load, and the second component represents the equivalent contribution of distributed generation. The physical constraint subnetwork includes a first constraint branch and a second constraint branch. The first constraint branch maps the output of the data-driven subnetwork to the power grid physical space and calculates the residual between the output and the power grid physical laws, generating a first physical loss term. The second constraint branch calculates the distributed generation based on the meteorological time-series data. The meteorological and physical upper limit of the power source output is determined, and a meteorological causal consistency constraint is applied to the second component based on the meteorological and physical upper limit to generate a second physical loss term. The physical constraint subnetwork is used to map the output of the data-driven subnetwork to the power grid physical space and calculate the residual between the output and the power grid physical law to generate a physical loss term. In the training and inference process, the physical information neural network model uses the static topology parameters of the power grid as the fixed input of the first constraint branch of the physical constraint subnetwork. The partially known distributed power source output time series data is used as partial labels, and combined with the first physical loss term and the second physical loss term, a joint loss function is constructed to train the physical information neural network model. The load forecasting module is used to input real-time measurement data into a trained physical information neural network model and output high-precision natural load time-series data after removing the influence of distributed power sources.

[0080] The functional explanations of each module in this embodiment are the same as those of a distribution network load decomposition prediction method based on physical information neural networks, and the technical effects are the same, so they will not be repeated here.

[0081] Those skilled in the art will recognize that the units of the various examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the invention.

[0082] In the embodiments provided by the present invention, it should be understood that the division of units is only a logical functional division. In actual implementation, there may be other division methods, such as multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored.

[0083] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0084] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the specification of the present invention.

Claims

1. A method for load decomposition and prediction of distribution networks based on physical information neural networks, characterized in that, Includes the following steps: S1. Acquire measurement data of each monitoring node in the target distribution network; the measurement data includes total injected power time series data, some known distributed power output time series data, meteorological time series data, and grid static topology parameters; the grid static topology parameters include node connection relationships and line impedance parameters; S2. Construct and train a physical information neural network model; the physical information neural network model includes a data-driven subnetwork and a physical constraint subnetwork; The data-driven subnetwork is used to decompose the total injected power time-series data into a first component and a second component; wherein the first component represents the actual natural load and the second component represents the equivalent contribution of distributed power sources. The physical constraint subnetwork includes a first constraint branch and a second constraint branch; the first constraint branch is used to map the output of the data-driven subnetwork to the power grid physical space, and calculate the residual between the output and the power grid physical law to generate a first physical loss term; the second constraint branch is used to calculate the meteorological physical upper limit of the distributed power generation output based on the meteorological time series data, and apply meteorological causal consistency constraints to the second component based on the meteorological physical upper limit to generate a second physical loss term. The physical constraint subnetwork is used to map the output of the data-driven subnetwork to the power grid physical space, and calculate the residual between the output and the power grid physical law to generate a physical loss term; In the training and inference process, the physical information neural network model uses the static topology parameters of the power grid as the fixed input of the first constraint branch of the physical constraint subnetwork; S3. Using the known distributed power output time-series data as partial labels, and combining the first physical loss term and the second physical loss term, a joint loss function is constructed to train the physical information neural network model. S4. Input the real-time measurement data into the trained physical information neural network model, and output the natural load time series data after removing the influence of distributed power sources.

2. The distribution network load decomposition and prediction method based on physical information neural network according to claim 1, characterized in that, The power grid physical laws include node power balance equations and branch power flow constraints; the first physical loss term is specifically the sum of the natural load of each node and the equivalent contribution of distributed generation output by the model. After power flow calculation using the static topology parameters of the power grid, the mean square error between this term and the actual total injected power is expressed as: in, This is the first physical loss term; This represents the total number of nodes in the distribution network. For nodes At any moment The actual measured value of the total injected power; Representation based on nodal admittance matrix The power flow calculation function is used to calculate the node. Theoretical injection power; For all nodes output by the model at time... The natural load vector; The output of the model at time 1 The equivalent contribution vector of distributed power sources; The node admittance matrix is ​​constructed from the node connection relationships and line impedance parameters in the static topology parameters of the power grid.

3. The distribution network load decomposition and prediction method based on physical information neural network according to claim 1, characterized in that, The data-driven sub-network adopts an encoder-decoder architecture; wherein, the encoder is used to extract deep features of the input time-series data, and the decoder includes two parallel output heads, which are used to regress the first component and the second component, respectively.

4. The distribution network load decomposition and prediction method based on physical information neural network according to claim 1, characterized in that, The specific method by which the second constraint branch generates the second physical loss term is as follows: Based on the solar irradiance, ambient temperature, and wind speed in the meteorological time-series data, the theoretical upper limit of the distributed power output of each node is calculated using a distributed power output physical conversion model. ; Second physical loss item Represented as: in, This represents the total number of nodes in the distribution network. The time window length for time series data; Nodes output by the model At any moment The equivalent contribution value of distributed power sources; For nodes At any moment The theoretical upper limit of output power of distributed power sources; This is a maximum value function used to extract the non-negative part of the value within parentheses.

5. The distribution network load decomposition and prediction method based on physical information neural network according to claim 1, characterized in that, The joint loss function is expressed as: in, For the joint loss function, This is the data fitting loss term, used to measure the error between the second component of the model output and the partially known distributed power output data. This refers to the first physical loss item; This is the second physical loss term; These are dynamically adjustable weighting coefficients.

6. The distribution network load decomposition and prediction method based on physical information neural network according to claim 5, characterized in that, The dynamically adjustable weighting coefficient Adaptive adjustments are made based on the missing rate of some known distributed power generation time-series data in the current training data. The adjustment rule is: the higher the missing rate, the higher the weight coefficient of the data fitting loss term. The smaller the value, the higher the weighting coefficient of the first physical loss term. and the weighting coefficient of the second physical loss term The larger the value, the higher the weighting coefficient becomes. When the missing rate exceeds a preset threshold, the weighting coefficient is updated according to the following rules: in, To preset the basic weight values, To set a preset missing rate threshold, The missing rate, The attenuation coefficient is... This is the growth coefficient.

7. The distribution network load decomposition and prediction method based on physical information neural network according to claim 1, characterized in that, After inputting real-time measurement data into the trained physical information neural network model and outputting natural load time-series data stripped of the influence of distributed power sources, the process further includes: Based on the aforementioned natural load time series data, dynamic monitoring of abnormal natural load growth is performed; When an anomaly is detected, the corresponding hidden layer features in the physical information neural network model are traced back to locate the main influencing nodes that caused the anomaly and the associated distributed power information.

8. The distribution network load decomposition and prediction method based on physical information neural network according to claim 1, characterized in that, The data-driven subnetwork adopts a parallel structure of temporal convolutional network and multi-head self-attention mechanism: The temporal convolutional network is used to extract local fluctuation features, and the multi-head self-attention mechanism is used to capture long-period dependencies; and the outputs of the temporal convolutional network and the multi-head self-attention mechanism are fused through an adaptive gating unit.

9. The distribution network load decomposition and prediction method based on physical information neural network according to claim 3, characterized in that, The decoder includes two parallel output heads, used for regressing the first component and the second component respectively: The head of the first component is output, and its output is modeled based on the Laplace distribution; wherein the location parameters of the Laplace distribution are obtained by linear transformation of the hidden layer output of the data-driven subnetwork. The header of the second component is output, and its output is modeled based on a Bernoulli-Gaussian mixture distribution; wherein, the mixture distribution includes a Bernoulli binary switching variable and a Gaussian distribution amplitude variable; the Bernoulli binary switching variable is used to characterize the intermittent operation state of the distributed power source, including zero or non-zero output, and the Gaussian distribution amplitude variable is used to regress the specific output amplitude in the non-zero state.

10. A distribution network load decomposition and prediction system based on physical information neural networks, characterized in that, include: The data acquisition module is used to acquire measurement data from each monitoring node of the target distribution network; the measurement data includes total injected power time series data, some known distributed power output time series data, meteorological time series data, and static topology parameters of the power grid. The model building and training module is used to build and train a physical information neural network model. The physical information neural network model includes a data-driven subnetwork and a physical constraint subnetwork. The data-driven subnetwork decomposes the total injected power time-series data into a first component and a second component. The first component represents the actual natural load, and the second component represents the equivalent contribution of distributed generation. The physical constraint subnetwork includes a first constraint branch and a second constraint branch. The first constraint branch maps the output of the data-driven subnetwork to the power grid physical space and calculates the residual between the output and the power grid physical laws, generating a first physical loss term. The second constraint branch calculates the distributed generation based on the meteorological time-series data. The meteorological and physical upper limit of the power source output is determined, and a meteorological causal consistency constraint is applied to the second component based on the meteorological and physical upper limit to generate a second physical loss term. The physical constraint subnetwork is used to map the output of the data-driven subnetwork to the power grid physical space and calculate the residual between the output and the power grid physical law to generate a physical loss term. In the training and inference process, the physical information neural network model uses the static topology parameters of the power grid as the fixed input of the first constraint branch of the physical constraint subnetwork. The partially known distributed power source output time series data is used as partial labels, and combined with the first physical loss term and the second physical loss term, a joint loss function is constructed to train the physical information neural network model. The load forecasting module is used to input real-time measurement data into a trained physical information neural network model and output high-precision natural load time-series data after removing the influence of distributed power sources.

Images