Real-time prediction method and system for power grid line loss rate PINN based on PDE dynamic following

Abstract

The present invention proposes a real-time prediction method and system for power line loss rate using a PDE that dynamically follows the PDE. This method first constructs a power line loss rate prediction model for the PINN neural network (PINN) that includes a BiGRU model and a dynamic PDE model. The method then divides historical power grid data into long-term, short-term, and real-time data sets. The long-term data set is then input into the prediction model for training to determine the network parameters of the BiGRU model, the weight ratio of the loss function, and the initial values ​​of the dynamic key parameters in the dynamic PDE model. The short-term data set is then input into the trained prediction model to obtain the loss function output value. If the output value is greater than the observation value, the dynamic key parameters are optimized and updated. Finally, the real-time data set is input into the updated prediction model to obtain the power line loss rate prediction result. By leveraging the observation-update function of the PDE model, the present invention enables the PINN neural network to dynamically follow changes in the power grid, thereby significantly improving the accuracy of real-time line loss rate prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of power grid control, and in particular relates to a method and system for real-time prediction of power grid line loss rate PINN with dynamic following of PDE. Background Art

[0002] Grid line loss refers to the energy lost during the transmission, transformation, and distribution stages of electricity. The line loss rate is typically defined as the ratio of line loss to total input power. The line loss rate is a key indicator for measuring grid operation quality, management, and economic efficiency. With the rapid economic development of my country, residential and industrial electricity consumption is increasing. Currently, line losses in low-voltage distribution networks account for approximately 40% of total grid losses, reducing the economic efficiency of grid operation. Therefore, reducing the line loss rate can effectively achieve low-carbon emissions, energy conservation, and loss reduction.

[0003] Accurately predicting the power grid line loss rate provides data support in the event of measurement failures. It also serves to verify the reliability and accuracy of measurement system data, promptly identifying abnormal measurement values ​​and faulty measurement devices. The predicted line loss rate data can be used to estimate future changes in power grid line losses, providing auxiliary indicators for technical upgrades, operational adjustments, and the implementation of energy-saving and loss-reduction measures. The predicted line loss rate results can also assist in power system planning and design, as well as network reconstruction, and are of great guiding significance for energy conservation and loss reduction, structural optimization, and improving the economic benefits of the power grid.

[0004] Because power grid line losses are influenced by a variety of factors, such as input power, output power, line length, temperature, load level, and grid operation mode, traditional power balance methods, which only predict line loss rates based on changes in input and output power, are difficult to accurately predict for complex power grids. Empirical formulas can estimate future line loss rates based on the historical load-loss relationship. However, in today's rapidly evolving power grid, this method struggles to accurately estimate future line loss rates. Using artificial intelligence neural networks, historical line loss data, load data, meteorological data, and other factors are used as input neurons. The trained neural network model can then predict future line loss rates.

[0005] In order to further improve the prediction accuracy of the neural network model for line loss rate and enhance the ability of the neural network model to process time series data, a recurrent unit can usually be introduced to form a recurrent neural network (RNN). Common ones include bidirectional recurrent neural networks (Bi-RNN) and long short-term memory networks (LSTM). The Chinese patent document with publication number CN110659779 B discloses a distribution system network loss prediction method based on long short-term memory network. The Chinese patent document with publication number CN 110598854 A discloses a substation line loss rate prediction method based on GRU model. This shows that the use of RNN model can be used to predict the line loss rate of the power grid.

[0006] These methods are based on purely data-driven neural network prediction methods, which rely heavily on the quantity and accuracy of historical data. This makes them prone to overfitting when predicting power grid line loss rates. Using historical data to predict future line loss rates is only applicable to long-term, fixed power grids. These purely data-driven neural network prediction methods have poor generalization capabilities and cannot predict the impact of grid structure changes on future line loss rates, making them difficult to provide reference forecast data for power grid upgrades.

[0007] In other forecasting areas, physical information model neural networks (PINNs) are often used to improve the poor generalization capabilities of purely data-driven neural network forecasting methods and reduce data dependence. However, the complex circuit topology of power grids makes it difficult to establish a suitable physical model, which is a problem faced by PINNs in the field of power line loss rate forecasting. Secondly, in traditional PINN forecasting, the physical processes or physical models are assumed to remain unchanged. However, power grid topology is constantly changing. New lines, line reconfigurations, and line aging can significantly alter the physical properties and structure of the power grid. Therefore, when using existing PINNs for real-time power line loss rate forecasting, they cannot accurately describe the current physical properties and structure of the power grid, resulting in a serious lack of real-time prediction accuracy. This makes it impossible to provide timely and effective reference data for power grid transactions. Summary of the Invention

[0008] The purpose of the present invention is to address the above-mentioned problems existing in the prior art and to provide a method and system for real-time prediction of power grid line loss rate PINN with dynamic following of PDE.

[0009] To achieve the above objectives, the technical solutions of the present invention are as follows:

[0010] In a first aspect, the present invention proposes a real-time prediction method for power line loss rate PINN with dynamic PDE following, comprising:

[0011] S1. Build a power grid PINN line loss rate prediction model that includes a BiGRU model and a dynamic PDE model; obtain historical data of the power grid and divide the historical data into long-term data sets, short-term data sets, and real-time data sets based on the period N days ago, 1-N days, and within 1 day, respectively;

[0012] S2. Input the long-term data set into the constructed prediction model for training to determine the network parameters of the BiGRU model, the weight ratio parameters of the loss function, and the initial values ​​of the dynamic key parameters in the dynamic PDE model;

[0013] S3. Input the recent data set into the trained prediction model to obtain the output value m of the loss function;

[0014] S4. Compare the output value m with its observation value m0. If m>m0, optimize the dynamic key parameters in the dynamic PDE model until m≤m0, and update the dynamic key parameters before entering S5. If m≤m0, directly enter S5.

[0015] S5. Input the real-time data set into the updated prediction model to obtain the power grid line loss rate prediction result.

[0016] In S1, the dynamic PDE model is:

[0017]

[0018]

[0019]

[0020] In the above formula, L (l，t) 、C (l，t) 、R ( l ，t) , G (l，t) are dynamic key parameters, representing distributed inductance, distributed capacitance, distributed resistance, and distributed conductance, respectively. l and t are the total length and time of the grid line, respectively. I and U are the current and voltage, respectively.

[0021] The S2 includes:

[0022] S21. Use algorithms to optimize the network parameters of the BiGRU model;

[0023] S22, based on the BiGRU model after algorithm optimization, adjust the weight ratio parameter of the loss function with the goal of minimizing the loss function, where the loss function for:

[0024]

[0025]

[0026]

[0027]

[0028] In the above formula, are respectively prediction error and physical error, λ data ,λ PDE are the predicted weight ratio and physical weight ratio respectively, μ^ is the true value of the power grid line loss rate, μ d is the line loss rate prediction value of the BiGRU model, μ p is the line loss rate value calculated by physical formula, t0, t x are the start and end time respectively, U0 and I0 are the voltage and current at the network end respectively, U l , I l are the voltage and current at the outgoing network end respectively;

[0029] S23. Based on the BiGRU model after the weight ratio parameters are adjusted, the dynamic key parameters in the dynamic PDE model are adjusted with the goal of minimizing the loss function to obtain initial values ​​of the dynamic key parameters;

[0030] S24. Determine whether the loss function has reached a minimum. If not, return to S22 and optimize the network parameters again until the loss function is minimized.

[0031] In S3, the output value m is calculated according to the following formula:

[0032]

[0033] In the above formula, λ data ,λ PDE are the predicted weight ratio and the physical weight ratio respectively, is the true value of the power line loss rate calculated from the i-th data point in the recent data set, are the power line loss rate values ​​of the i-th data point in the recent data set predicted by the BiGRU network and calculated by the physical formula, respectively, and n is the number of data points in the recent data set;

[0034] In S4, the inspection value m0 is calculated according to the following formula:

[0035]

[0036] In the above formula, δ is the tolerance factor with a value range of [0, 2], λ data ,λ PED are the prediction weight ratio and physical weight ratio respectively, is the true value of the power line loss rate calculated from the i-th data point in the long-term data set, are the power grid line loss rate values ​​of the i-th data point in the recent data set predicted by the BiGRU network and calculated by the physical formula, respectively, and k is the number of data points in the long-term data set.

[0037] The S21 uses the Bayesian optimization algorithm to optimize the network parameters of the BiGRU model, including the number of hidden units, the maximum number of training rounds, the initial learning rate, the L2 regularization parameter, the gradient threshold, the Dropout rate, and the learning rate adjustment factor;

[0038] The S4 uses a particle swarm optimization algorithm to optimize the dynamic key parameters in the dynamic PDE model.

[0039] In the second aspect, the present invention proposes a PDE dynamic following power grid line loss rate PINN real-time prediction system, which includes a prediction model construction module, a data partitioning module, a prediction model training module, a prediction model inspection module, a dynamic parameter update module, and a real-time prediction module;

[0040] The prediction model building module is used to build a power grid PINN line loss rate prediction model including a BiGRU model and a dynamic PDE model;

[0041] The data partitioning module is used to obtain historical data of the power grid and divide the historical data into long-term data sets, short-term data sets and real-time data sets according to N days ago, 1-N days and within 1 day respectively;

[0042] The prediction model training module is used to input the long-term data set into the constructed prediction model for training to determine the network parameters of the BiGRU model, the weight ratio parameters of the loss function, and the initial values ​​of the dynamic key parameters in the dynamic PDE model;

[0043] The prediction model inspection module is used to input the recent data set into the trained prediction model, obtain the output value m of the loss function, and compare the output value m with the inspection value m0. If m>m0, the dynamic parameter update module is started;

[0044] The dynamic parameter updating module is used to optimize the dynamic key parameters in the dynamic PDE model until m≤m0, and update the dynamic key parameters;

[0045] The real-time prediction module is used to input the real-time data set into the updated prediction model to obtain the power grid line loss rate prediction result.

[0046] The prediction model building module is used to build the following dynamic PDE model:

[0047]

[0048]

[0049]

[0050] In the above formula, L (l，t) 、C (l，t) 、R (l，t) , G (l，t) are dynamic key parameters, representing distributed inductance, distributed capacitance, distributed resistance, and distributed conductance, respectively. l and t are the total length and time of the grid line, respectively. I and U are the current and voltage, respectively.

[0051] The prediction model training module includes a network parameter optimization unit, a weight ratio parameter adjustment unit, a dynamic key parameter adjustment unit, and a judgment unit;

[0052] The network parameter optimization unit is used to optimize the network parameters of the BiGRU model using an algorithm;

[0053] The weight ratio parameter adjustment unit is used to adjust the weight ratio parameter of the loss function based on the BiGRU model optimized by the algorithm with the goal of minimizing the loss function, wherein the loss function for:

[0054]

[0055]

[0056]

[0057]

[0058] In the above formula, are respectively prediction error and physical error, λ data ,λ PDE are the predicted weight ratio and physical weight ratio respectively, μ^ is the true value of the power grid line loss rate, μ d is the line loss rate prediction value of the BiGRU model, μ p is the line loss rate value calculated by physical formula, t0, t x are the start and end time respectively, U0 and I0 are the voltage and current at the network end respectively, U l , I l are the voltage and current at the outgoing network end respectively;

[0059] The dynamic key parameter adjustment unit is used to adjust the dynamic key parameters in the dynamic PDE model based on the BiGRU model after the weight ratio parameter is adjusted, with the loss function minimized as the goal, to obtain the initial value of the dynamic key parameter;

[0060] The judgment unit is used to judge whether the loss function reaches the minimum. If not, the network parameter optimization unit is started to optimize the network parameters again until the loss function reaches the minimum.

[0061] The prediction model inspection module calculates the output value m according to the following formula

[0062]

[0063] In the above formula, λ data ,λ PDE are the prediction weight ratio and physical weight ratio respectively, is the true value of the power line loss rate calculated from the i-th data point in the recent data set, are the power line loss rate values ​​of the i-th data point in the recent data set predicted by the BiGRU network and calculated by the physical formula, respectively, and n is the number of data points in the recent data set;

[0064] The investigation value m0 is calculated according to the following formula:

[0065]

[0066] In the above formula, δ is the tolerance factor with a value range of [0, 2], λ data ,λ PDE are the prediction weight ratio and physical weight ratio respectively, is the true value of the power line loss rate calculated from the i-th data point in the long-term data set, are the power grid line loss rate values ​​of the i-th data point in the recent data set predicted by the BiGRU network and calculated by the physical formula, respectively, and k is the number of data points in the long-term data set.

[0067] The network parameter optimization unit uses the Bayesian optimization algorithm to optimize the network parameters of the BiGRU model, including the number of hidden units, the maximum number of training rounds, the initial learning rate, the L2 regularization parameter, the gradient threshold, the Dropout rate, and the learning rate adjustment factor;

[0068] The dynamic parameter updating module uses a particle swarm optimization algorithm to optimize the dynamic key parameters in the dynamic PDE model.

[0069] Compared with the prior art, the present invention has the following beneficial effects:

[0070] The present invention provides a PDE dynamically following power grid line loss rate PINN real-time prediction method, which first constructs a power grid PINN line loss rate prediction model including a BiGRU model and a dynamic PDE model, and obtains historical data of the power grid, and divides the historical data into a long-term data set, a short-term data set and a real-time data set according to N days ago, 1-N days and within 1 day, respectively. The long-term data set is then input into the constructed prediction model for training to determine the network parameters of the BiGRU model, the weight ratio parameters of the loss function and the initial values ​​of the dynamic key parameters in the dynamic PDE model. The short-term data set is then input into the trained prediction model to obtain the output value m of the loss function, and the output value m is compared with its observation value m0. If m>m0, the dynamic key parameters in the dynamic PDE model are optimized until m≤m0, and the dynamic key parameters are updated. Finally, the real-time data set is input into the updated prediction model to obtain a power grid line loss rate prediction result. This method introduces a dynamic PDE model with adjustable key parameters and establishes a dynamic update mode of the PDE model, thereby realizing the inspection-update function of the PDE model. On the one hand, this inspection-update function helps the PINN neural network dynamically follow the latest changes in the power grid, thereby greatly improving the accuracy of real-time prediction of the power grid line loss rate; on the other hand, the PDE model update step will only be started when the power grid structure has undergone major changes and exceeded the inspection value, avoiding the PINN model from calculating and iterating the network structure, loss function weights and key parameters every time it runs, greatly saving the computing power of the intelligent server; at the same time, this method can be used for real-time changes in the line loss rate after power grid transformation, providing timely effect evaluation for power grid transformation, without waiting for the power grid to operate for a long time again and obtaining a large amount of data before conducting data analysis and effect evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 This is a flowchart of the method described in Example 1.

[0072] Figure 2 This is a schematic diagram of the 10kV distribution circuit structure in Example 1.

[0073] Figure 3 Schematic diagram of the BiGRU model in Example 1.

[0074] Figure 4 This is a flow chart of the Bayesian optimization algorithm in Example 1.

[0075] Figure 5 Schematic diagram of forecast error on weekdays.

[0076] Figure 6 Schematic diagram of forecast error on non-working days.

[0077] Figure 7 2 is a structural diagram of the system of the present invention. DETAILED DESCRIPTION

[0078] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0079] Example 1:

[0080] This embodiment takes the actual 10kV distribution circuit structure of a township power grid area (including a 110 / 10kV substation with a capacity of 2*100MVA, a total of 39 substations in the planning area, including 58 10kV lines) as the research object, and implements the power line loss rate PINN real-time prediction method of the present invention with a PDE dynamic following method. Figure 1 The specific steps are as follows:

[0081] 1. Collect data on the grid's input and sales electricity. Perform 24-hour predictive sampling of the distribution network's losses, input voltage U0, and input current I0. The data is sampled every 20 minutes, with a total of 72 sampling points throughout the day. Obtain one year of historical data and group data from 30 days ago into the long-term dataset, data from 1 to 30 days into the near-term dataset, and data from less than 1 day into the real-time dataset.

[0082] 2. Preprocess the above data, including supplementing blank values, correcting abnormal data, normalizing the data, and constructing the time series of corresponding feature quantities. The missing value processing formula is:

[0083]

[0084] In the above formula, D(X) is the missing value, n is the length of the time series data containing the missing, and Y i is the function value of the known data point, L i (x) is the Lagrangian basis function;

[0085] In order to avoid the problem of large numerical differences, the Min-Max Normalization method is used to normalize the input power and sales power data and construct the time series of the corresponding feature quantities.

[0086] 3. Construct a PINN line loss rate prediction model for the power grid, including:

[0087] 1) Use the Tensorflow library to build a BiGRU model in Python, such as Figure 2 As shown. The model sets 4 layers of BiGRU with 128 units. In the initial Bi-GRU network model, the output information at a certain time t is the sum of the outputs of the forward hidden layer and the backward hidden layer. The specific calculation formula is as follows:

[0088]

[0089] In the above formula, GRU(·) is a gated recurrent unit, are the outputs of the forward and backward hidden layers respectively, is the hidden state of the forward hidden state at the previous moment, is the hidden state of the backward hidden state at the next moment, w t 、v t are the weights corresponding to the forward hidden layer state and the backward hidden layer state of Bi-GRU at time t, respectively, t is the bias value corresponding to the hidden layer state at time t.

[0090] 2) Constructing a dynamic PDE model

[0091] In power grid distribution systems, the number of transmission lines is enormous, the connections are complex, and the number of components, such as transformers, is difficult to estimate. To simplify the physical structure of the power grid, this embodiment simplifies the impact of these factors on the distribution network's transmission lines into the effects of unit inductance, resistance, capacitance, and conductance on a transmission line of length l. This simplifies the power grid transmission line circuit into a single transmission line circuit. The following dynamic PDE model is obtained:

[0092]

[0093]

[0094]

[0095] In the above formula, L (l，t) 、C (l，t) 、R (l，t) , G (l，t) are dynamic key parameters, representing distributed inductance, distributed capacitance, distributed resistance, and distributed conductance, respectively. They are all functions that change with the total length l of the grid line (determined according to the total length of the township power grid) and time t. I and U are current and voltage, respectively.

[0096] The BiGRU model and the dynamic PDE model constitute the power grid PINN line loss rate prediction model.

[0097] 4. The Bayesian optimization algorithm (BO algorithm) is used to optimize the network parameters of the BiGRU model, including the number of hidden units, the maximum number of training rounds, the initial learning rate, the L2 regularization parameter, the gradient threshold, the dropout rate, and the learning rate adjustment factor.

[0098] The BO algorithm maximizes or minimizes a black box objective function based on probability models such as Bayesian theorem and Gaussian process. It continuously selects the next sampling point in the search space and then estimates the optimal value of the function based on the existing sampled data, gradually converging to the global optimal solution. Its expression is:

[0099]

[0100] In the above formula, f(h) is the prior distribution model, h * is the optimal parameter value of the constraint domain of f(h), and H is the candidate set.

[0101] Compared with grid search and random search, Bayesian optimization algorithm can achieve satisfactory optimization results with fewer iterations. The specific workflow of BO algorithm is as follows: Figure 3 shown.

[0102] The optimized network parameters in this embodiment are: the number of hidden layer neurons is 30, the time window size is 50, the learning rate is 0.001, and the drop rate is 0.1.

[0103] 5. Based on the BiGRU model optimized by the BO algorithm, the weight ratio parameters of the loss function are adjusted with the goal of minimizing the loss function. Among them, the loss function for:

[0104]

[0105]

[0106]

[0107]

[0108] μ^=(W in -W out ) / W in

[0109] In the above formula, are respectively prediction error and physical error, λ data ,λ PDE are the predicted weight ratio and physical weight ratio respectively, μ^ is the true value of the power grid line loss rate, μ d is the line loss rate prediction value of the BiGRU model, μ p is the line loss rate value calculated by physical formula, t0, t x are the start and end time respectively, U0 and I0 are the voltage and current at the network end respectively, U l , I l are the voltage and current at the outgoing network end, W in 、W outThey are respectively the input electricity and the sold electricity.

[0110] The adjusted weight ratio parameter obtained in this embodiment is: data =0.7,λ PDE =0.3.

[0111] 6. Based on the BiGRU model after weight ratio parameter adjustment, the dynamic key parameters in the dynamic PDE model are adjusted with the goal of minimizing the loss function to obtain the initial values ​​of the dynamic key parameters.

[0112] The initial value of the dynamic key parameter obtained in this embodiment is: L (l，t) =10mH / km, C (l，t) =600nF / km, R (l，t) =2Ω / km, G (l，t) =10 -6 S / km.

[0113] 7. Determine whether the loss function has reached the minimum. If not, return to step 4 and optimize the network parameters again until the loss function is minimized.

[0114] 8. Enter the PDE equation dynamic update module, input the recent data set into the trained prediction model, and calculate the output value m of the loss function according to the following formula, where the formula is as follows:

[0115]

[0116] In the above formula, λ data ,λ PDE are the prediction weight ratio and physical weight ratio respectively, is the true value of the power line loss rate calculated from the i-th data point in the recent data set, are the power line loss rate values ​​of the i-th data point in the recent data set predicted by the BiGRU network and calculated by the physical formula, respectively, and n is the number of data points in the recent data set;

[0117] The output value m obtained in this embodiment is 0.82.

[0118] 9. Calculate the inspection value m0 according to the following formula:

[0119]

[0120] In the above formula, δ is the tolerance factor with a value range of [0, 2], which represents the tolerance level of the output value m, and λ data ,λ PDE are the prediction weight ratio and physical weight ratio respectively, is the true value of the power line loss rate calculated from the i-th data point in the long-term data set, are the power grid line loss rate values ​​of the i-th data point in the recent data set predicted by the BiGRU network and calculated by the physical formula, respectively, and k is the number of data points in the long-term data set.

[0121] The investigated value m0 obtained in this embodiment is 0.5.

[0122] 10. Compare the output value m with its observation value m0. If m≤m0, it indicates that the dynamic key parameters set based on the long-term data set still meet the expectations for the boundary restriction effect on the recent data set, and the aforementioned prediction model can be directly used for real-time prediction; if m>m0, it indicates that the dynamic key parameters set based on the long-term data set cannot meet the expectations for the boundary restriction effect on the recent data set, that is, the recent changes in the power grid make the PDE model no longer able to accurately reflect the physical information of the current power grid, and the dynamic key parameters need to be reset. At this time, the particle swarm optimization algorithm is used to optimize the dynamic key parameters in the dynamic PDE model until m≤m0, and the dynamic key parameters are updated.

[0123] The goal of Particle Swarm Optimization (PSO) is to find the global optimal solution of the fitness function in D-dimensional space, that is, the optimal parameter combination. The position x and velocity v of each particle in the particle swarm describe its behavior in space. The position of the particle is a 4-dimensional vector,

[0124] x={x1,x2,x3,x4}={L (l，t) , C (l，t) , R (l，t) , G (l，t)}

[0125] During the search process, the particle swarm tracks the local optimal solution P best and the global optimal solution G best , update the position x and velocity v. Its expression is:

[0126] v i (t+1)=ω×v i (t)+c1×r1×(P best_id -x i (t))+c2×r2×(G best_id -x i (t))

[0127] x i (t+1=x i (t)+v i (t+1)

[0128] In the above formula, x i (t), v i (t) are the current position and current speed, P best_idis the local optimal solution of the i-th particle, G best_id is the global optimal solution of the i-th particle, c1 and c2 are acceleration constants, r1 and r2 are random numbers, and ω is the inertia weight, which affects the convergence of the PSO algorithm. The larger the ω value, the weaker the particle's local optimization ability and the stronger its global search ability, and vice versa. Therefore, the ω value is set to a large value at the beginning of the iteration and gradually reduced during the iteration process. Its attenuation formula is:

[0129]

[0130] In the above formula, ω max 、ω min are the maximum and minimum values ​​of the inertia weight respectively, n is the current number of iterations, n max is the total number of iterations.

[0131] In this embodiment, m>m0, so the particle swarm optimization algorithm is used to optimize the dynamic key parameters. After optimization, m=0.33 is calculated, which satisfies m≤m0. The dynamic key parameters are updated to: L (l，t) =8.3mH / km, C (l，t) =657nF / km, R (l，t) =3.4Ω / km, G (l，t) =10.26 -6 S / km.

[0132] 11. Enter the real-time prediction module and input the real-time data set into the prediction model after updating the dynamic key parameters to obtain the prediction results of the power line loss rate on working days and non-working days, respectively. Figure 5 、 Figure 6 shown.

[0133] It can be seen that the line loss rate prediction value obtained by this method is very close to the changing pattern of the true value.

[0134] Example 2:

[0135] A real-time prediction of power line loss rate PINN with dynamic following of PDE, such as Figure 7 As shown, it includes a prediction model construction module, a data partitioning module, a prediction model training module, a prediction model inspection module, a dynamic parameter updating module, and a real-time prediction module.

[0136] The prediction model building module is used to build a power grid PINN line loss rate prediction model including a BiGRU model and a dynamic PDE model, wherein the dynamic PDE model is:

[0137]

[0138]

[0139]

[0140] In the above formula, L (l，t) 、C (l，t) 、R (l，t) , G (l，t) are dynamic key parameters, representing distributed inductance, distributed capacitance, distributed resistance, and distributed conductance, respectively. l and t are the total length and time of the grid line, respectively. I and U are the current and voltage, respectively.

[0141] The data partitioning module is used to obtain historical data of the power grid, and divide the historical data into a long-term data set, a short-term data set and a real-time data set according to N days ago, 1-N days and within 1 day respectively.

[0142] The prediction model training module is used to input the long-term data set into the constructed prediction model for training to determine the network parameters of the BiGRU model, the weight ratio parameters of the loss function, and the initial values ​​of the dynamic key parameters in the dynamic PDE model, including a network parameter optimization unit, a weight ratio parameter adjustment unit, a dynamic key parameter adjustment unit, and a judgment unit;

[0143] The network parameter optimization unit is used to optimize the network parameters of the BiGRU model using a Bayesian optimization algorithm, including the number of hidden units, the maximum number of training rounds, the initial learning rate, the L2 regularization parameter, the gradient threshold, the Dropout rate, and the learning rate adjustment factor;

[0144] The weight ratio parameter adjustment unit is used to adjust the weight ratio parameter of the loss function based on the BiGRU model optimized by the algorithm with the goal of minimizing the loss function, wherein the loss function for:

[0145]

[0146]

[0147]

[0148]

[0149] In the above formula, are respectively prediction error and physical error, λ data ,λ PDE are the predicted weight ratio and physical weight ratio respectively, μ^ is the true value of the power grid line loss rate, μ d is the line loss rate prediction value of the BiGRU model, μ p is the line loss rate value calculated by physical formula, t0, t x are the start and end time respectively, U0 and I0 are the voltage and current at the network end respectively, Ul , I l are the voltage and current at the outgoing network end respectively;

[0150] The dynamic key parameter adjustment unit is used to adjust the dynamic key parameters in the dynamic PDE model based on the BiGRU model after the weight ratio parameter is adjusted, with the loss function minimized as the goal, to obtain the initial value of the dynamic key parameter;

[0151] The judgment unit is used to judge whether the loss function reaches the minimum. If not, the network parameter optimization unit is started to optimize the network parameters again until the loss function reaches the minimum.

[0152] The prediction model inspection module is used to input the recent data set into the trained prediction model to obtain the output value m of the loss function, and compare the output value m with the inspection value m0. If m>m0, the dynamic parameter update module is started to perform dynamic parameter update, wherein the output value m is calculated according to the following formula:

[0153]

[0154] In the above formula, λ data ,λ PDE are the prediction weight ratio and physical weight ratio respectively, is the true value of the power line loss rate calculated from the i-th data point in the recent data set, are the power line loss rate values ​​of the i-th data point in the recent data set predicted by the BiGRU network and calculated by the physical formula, respectively, and n is the number of data points in the recent data set;

[0155] The investigation value m0 is calculated according to the following formula:

[0156]

[0157] In the above formula, δ is the tolerance factor with a value range of [0, 2], λ data ,λ PDF are the prediction weight ratio and physical weight ratio respectively, is the true value of the power line loss rate calculated from the i-th data point in the long-term data set, are the power grid line loss rate values ​​of the i-th data point in the recent data set predicted by the BiGRU network and calculated by the physical formula, respectively, and k is the number of data points in the long-term data set.

[0158] The dynamic parameter updating module is used to optimize the dynamic key parameters in the dynamic PDE model using a particle swarm optimization algorithm until m≤m0, and update the dynamic key parameters.

[0159] The real-time prediction module is used to input the real-time data set into the updated prediction model to obtain the power grid line loss rate prediction result.

Claims

1. A real-time prediction method for power line loss rate PINN with PDE dynamic following, characterized by: The method comprises: S1. Build a power grid PINN line loss rate prediction model that includes a BiGRU model and a dynamic PDE model; obtain historical data of the power grid and divide the historical data into long-term data sets, short-term data sets, and real-time data sets based on the period N days ago, 1-N days, and within 1 day, respectively; S2. Input the long-term data set into the constructed prediction model for training to determine the network parameters of the BiGRU model, the weight ratio parameters of the loss function, and the initial values ​​of the dynamic key parameters in the dynamic PDE model, including: S21. Use algorithms to optimize the network parameters of the BiGRU model; S22, based on the BiGRU model after algorithm optimization, adjust the weight ratio parameter of the loss function with the goal of minimizing the loss function, where the loss function for: ； ； ； ； In the above formula, 、 are prediction error and physical error respectively, 、 are the prediction weight ratio and physical weight ratio respectively, is the true value of the power grid line loss rate, is the line loss rate prediction value of the BiGRU model, is the line loss rate value calculated by physical formula, 、 are the start and end time respectively, 、 are the voltage and current at the network end, 、 are the voltage and current at the outgoing network end respectively; S23. Based on the BiGRU model after the weight ratio parameters are adjusted, the dynamic key parameters in the dynamic PDE model are adjusted with the goal of minimizing the loss function to obtain initial values ​​of the dynamic key parameters; S24, determine whether the loss function reaches the minimum. If not, return to S22 and optimize the network parameters again until the loss function reaches the minimum. S3. Input the recent data set into the trained prediction model to obtain the output value of the loss function ; S4, compare output value Its inspection value ,like , then optimize the dynamic key parameters in the dynamic PDE model until , and update the dynamic key parameters and enter S5; if , then go directly to S5; S5. Input the real-time data set into the updated prediction model to obtain the power grid line loss rate prediction result.

2. The method for real-time prediction of power line loss rate PINN with dynamic PDE following according to claim 1 is characterized in that: In S1, the dynamic PDE model is: ； In the above formula, 、 、 、 are dynamic key parameters, representing distributed inductance, distributed capacitance, distributed resistance, and distributed conductance, respectively. 、 are the total length of the power grid line, time, I 、 U are current and voltage respectively.

3. A method for real-time prediction of power line loss rate PINN with dynamic PDE following according to claim 1 or 2, characterized in that: In S3, the output value Calculated according to the following formula: ； In the above formula, 、 are the prediction weight ratio and physical weight ratio respectively, To collect the latest data i The true value of the power grid line loss rate calculated from the data points is: 、 The first and second recent data sets are predicted by BiGRU network and calculated by physical formula. i The power grid line loss rate value of the data point, is the number of data points in the recent dataset; In S4, the inspection value Calculated according to the following formula: ； In the above formula, is a tolerance factor with a value range of [0, 2], 、 are the prediction weight ratio and physical weight ratio respectively, To collect the first i The true value of the power grid line loss rate calculated from the data points is: 、 The first and second recent data sets are predicted by BiGRU network and calculated by physical formula. i The power grid line loss rate value of the data point, is the number of data points in the long-term dataset.

4. The method for real-time prediction of power line loss rate PINN with dynamic PDE following according to claim 1 is characterized in that: The S21 uses the Bayesian optimization algorithm to optimize the network parameters of the BiGRU model, including the number of hidden units, the maximum number of training rounds, the initial learning rate, the L2 regularization parameter, the gradient threshold, the Dropout rate, and the learning rate adjustment factor; The S4 uses a particle swarm optimization algorithm to optimize the dynamic key parameters in the dynamic PDE model.

5. A PDE dynamic following power grid line loss rate PINN real-time prediction system, characterized by: The system includes a prediction model construction module, a data partitioning module, a prediction model training module, a prediction model inspection module, a dynamic parameter updating module, and a real-time prediction module; The prediction model building module is used to build a power grid PINN line loss rate prediction model including a BiGRU model and a dynamic PDE model; The data partitioning module is used to obtain historical data of the power grid and divide the historical data into long-term data sets, short-term data sets and real-time data sets according to N days ago, 1-N days and within 1 day respectively; The prediction model training module is used to input the long-term data set into the constructed prediction model for training to determine the network parameters of the BiGRU model, the weight ratio parameters of the loss function, and the initial values ​​of the dynamic key parameters in the dynamic PDE model, including a network parameter optimization unit, a weight ratio parameter adjustment unit, a dynamic key parameter adjustment unit, and a judgment unit; The network parameter optimization unit is used to optimize the network parameters of the BiGRU model using an algorithm; The weight ratio parameter adjustment unit is used to adjust the weight ratio parameter of the loss function based on the BiGRU model optimized by the algorithm with the goal of minimizing the loss function, wherein the loss function for: ； ； ； ； In the above formula, 、 are prediction error and physical error respectively, 、 are the prediction weight ratio and physical weight ratio respectively, is the true value of the power grid line loss rate, is the line loss rate prediction value of the BiGRU model, is the line loss rate value calculated by physical formula, 、 are the start and end time respectively, 、 are the voltage and current at the network end, 、 are the voltage and current at the outgoing network end respectively; The dynamic key parameter adjustment unit is used to adjust the dynamic key parameters in the dynamic PDE model based on the BiGRU model after the weight ratio parameter is adjusted, with the loss function minimized as the goal, to obtain the initial value of the dynamic key parameter; The judgment unit is used to judge whether the loss function has reached a minimum. If not, the network parameter optimization unit is started to optimize the network parameters again until the loss function is minimized. The prediction model inspection module is used to input the recent data set into the trained prediction model to obtain the output value of the loss function , and output the value and inspection value Compare, if , then start the dynamic parameter update module; The dynamic parameter update module is used to optimize the dynamic key parameters in the dynamic PDE model until , and update dynamic key parameters; The real-time prediction module is used to input the real-time data set into the updated prediction model to obtain the power grid line loss rate prediction result.

6. The PDE dynamic following PINN real-time prediction system for power line loss rate according to claim 5 is characterized in that: The prediction model building module is used to build the following dynamic PDE model: ； In the above formula, 、 、 、 are dynamic key parameters, representing distributed inductance, distributed capacitance, distributed resistance, and distributed conductance, respectively. 、 are the total length of the power grid line, time, I 、 U are current and voltage respectively.

7. A PDE dynamic following power grid line loss rate PINN real-time prediction system according to claim 5 or 6, characterized in that: The prediction model inspection module calculates the output value according to the following formula : ； In the above formula, 、 are the prediction weight ratio and physical weight ratio respectively, To collect the latest data i The true value of the power grid line loss rate calculated from the data points is: 、 The first and second recent data sets are predicted by BiGRU network and calculated by physical formula. i The power grid line loss rate value of the data point, is the number of data points in the recent dataset; The investigation value Calculated according to the following formula: ； In the above formula, is a tolerance factor with a value range of [0, 2], 、 are the prediction weight ratio and physical weight ratio respectively, To collect the first i The true value of the power grid line loss rate calculated from the data points is: 、 The first and second recent data sets are predicted by BiGRU network and calculated by physical formula. i The power grid line loss rate value of the data point, is the number of data points in the long-term dataset.

8. The PDE dynamic following PINN real-time prediction system for power line loss rate according to claim 5 is characterized in that: The network parameter optimization unit uses the Bayesian optimization algorithm to optimize the network parameters of the BiGRU model, including the number of hidden units, the maximum number of training rounds, the initial learning rate, the L2 regularization parameter, the gradient threshold, the Dropout rate, and the learning rate adjustment factor; The dynamic parameter updating module uses a particle swarm optimization algorithm to optimize the dynamic key parameters in the dynamic PDE model.

Images