Substation load prediction method and system based on physical model pre-training

Abstract

The invention discloses a transformer station load prediction method and system based on physical model pre-training, and the method comprises the steps: carrying out the pre-training of a load prediction model through simulation data outputted by a target transformer station physical model, and carrying out the rolling optimization prediction of the load prediction model through combining actual real-time data. Therefore, the model can better adapt to the real environment while keeping the adaptive capacity to unknown data, so that the generalization ability of the model is improved; the simulation data and the actual data can provide different gradient information, by combining the gradient information of the simulation data and the actual data, the parameters of the model can be adjusted more effectively, the performance of the model is improved, the model can learn more load characteristics and laws, and therefore the prediction accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of load forecasting, and in particular relates to a method and system for substation load forecasting based on physical model pre-training. Background Art

[0002] As global energy demand continues to grow, it is expected that by 2050, global electricity demand will increase by nearly 50% from the current level, which puts higher requirements on the efficiency and sustainability of the energy system. At the same time, the continuous advancement of energy conservation and emission reduction policies, especially under the goals of the Paris Agreement, countries have pledged to reduce carbon emissions, making the power industry a key area of ​​transformation. As the indispensable "heart" of the power system, the energy consumption of substations accounts for a considerable proportion of the total energy consumption of the power grid. According to the International Energy Agency, the energy consumption of substations can account for 1%-3% of the transmitted power, and this proportion is even higher in old or inefficient substations.

[0003] With the continuous growth of global energy demand and the advancement of energy conservation and emission reduction policies, the energy consumption management and optimization of substations are becoming increasingly important. The existing load forecasting uses a single data source, which only inputs historical experience data, such as equipment rated power, operating time, etc., but historical experience data are usually lost and missing, and there is a lack of data to fully train the prediction model; the existing load forecasting uses a small number of data sources, which only considers static energy efficiency indicators and ignores the dynamic characteristics of load changes over time, weather, economic activities and other factors. According to research, the actual load fluctuation of the substation can reach 20%-50% of the rated load. It is difficult to accurately reflect the actual energy consumption status and load change trend of the substation by using only single-dimensional data for prediction. Summary of the invention

[0004] In order to solve the problems existing in the prior art, the present invention proposes a substation load prediction method and system based on physical model pre-training.

[0005] The technical solution of the present invention is as follows:

[0006] A substation load forecasting method based on physical model pre-training, comprising:

[0007] Based on the physical model of the target substation, simulate and generate the corresponding substation unit area building load data under different meteorological conditions and heat dissipation conditions, and construct a data set based on the meteorological conditions and the corresponding substation unit area building load data;

[0008] Pre-training multiple load forecasting models based on the data set;

[0009] Perform load forecasting based on multiple pre-trained load forecasting models, evaluate the load forecasting results, optimize the forecasting model based on the evaluation results, and obtain the final load forecasting model;

[0010] The final load forecasting model is subjected to rolling optimization forecasting based on the actual real-time data of the target substation.

[0011] Furthermore, the simulation to generate the corresponding substation unit area building load data under different meteorological condition data and heat dissipation condition data is performed using EnergyPlus energy consumption simulation software and Dest energy consumption simulation software.

[0012] Furthermore, the meteorological condition data includes temperature condition data, humidity condition data, air wet-bulb temperature condition data and wind speed condition data; the heat dissipation condition data is the heat generated by the equipment per unit time; the substation unit area building load data includes the electricity consumption per unit area and the air conditioning energy consumption.

[0013] Furthermore, the specific method of pre-training multiple load forecasting models based on the data set is to perform sub-step 1 and sub-step 2 on multiple load forecasting models respectively; wherein,

[0014] Sub-step 1:

[0015] Dataset division: Sort the substation unit area building load data corresponding to different meteorological conditions generated by simulation according to time series correlation, select the first 90% of continuous data as the training set for model training; the remaining 10% of the data is used as the test set to verify the model performance;

[0016] Data normalization: The specific calculation formula for normalization is as follows:

[0017]

[0018] Among them, X norm is the normalized value, Xmin and Xmax are the minimum and maximum values ​​of the data respectively;

[0019] Sub-step 2:

[0020] The load forecasting model is pre-trained using the meteorological condition data in the training set as the model input data and the substation unit area building load data as the model output data;

[0021] The test set is used to verify the performance of the pre-trained load forecasting model.

[0022] Furthermore, the load forecasting model includes a Transformer model, an LSTM neural network, a support vector machine, and an RBF neural network; wherein the Transformer model is based on a seq2seq architecture and consists of an encoder and a decoder, wherein the encoder converts an input sequence into a series of continuous vector representations, and the decoder generates a target sequence according to the vector representations;

[0023] The LSTM neural network is composed of a plurality of stacked LSTM layers and one or more fully connected layers, and the number of units in each LSTM layer is set;

[0024] The RBF neural network consists of an input layer, a hidden layer and an output layer. The hidden layer neurons use the radial basis function as an activation function. The input layer receives input data. Each hidden layer neuron uses the radial basis function to calculate the distance between the input data and the center point, and generates an output accordingly. The formula of the Gaussian function as the radial basis function can be expressed as:

[0025]

[0026] Among them, x is the input data, c i is the center point of the i-th hidden layer neuron, σ i is the width parameter of the neuron, and ∥·∥ represents the Euclidean distance.

[0027] In the output layer, the output of the hidden layer is linearly combined to obtain the final prediction result. The calculation formula of the output layer can be expressed as:

[0028]

[0029] Among them, y is the prediction result, N is the number of neurons in the hidden layer, and w i is the weight from the i-th hidden layer neuron to the output layer, and b is the bias term;

[0030] The optimization problem of the support vector machine is expressed as:

[0031]

[0032] Among them, w is the weight vector, b is the bias term, C is the regularization parameter used to control the balance between model complexity and training error, and ξi and ξi* are slack variables used to deal with noise and outliers in the data.

[0033] By solving the optimization problem, the weight vector w and the bias term b are obtained, and then the prediction function f(x)=w·x+b is obtained.

[0034] Furthermore, the evaluation of the load forecasting results and the optimization of the forecasting model based on the evaluation results to obtain the final load forecasting model include:

[0035] The mean square error, mean absolute error and R2 score are used to evaluate the load forecasting results of different load forecasting models, and then the model with the best evaluation result is selected as the final load forecasting model according to actual needs, or multiple load forecasting models are combined for prediction and weights are assigned according to the load forecasting results.

[0036] Furthermore, the specific method of performing rolling optimization prediction on the final load forecasting model based on the actual real-time data of the target substation includes the following sub-steps 1 and 2; wherein,

[0037] Sub-step 1: Use a forecast formula in the rolling forecast to represent the behavior of the model in each forecast. The forecast formula uses a moving average formula or an exponential smoothing formula.

[0038] Sub-step 2: Using k-fold cross validation, the average performance index of the model is calculated by the following formula:

[0039]

[0040] Where k is the number of subsets, n i is the number of samples in the i-th test set, Y ij is the actual value of the jth sample in the i-th test set, is the corresponding predicted value. This formula calculates the average RMSE on all test sets as the final evaluation indicator of model performance.

[0041] A substation load forecasting system based on physical model pre-training, including a data set building module, a pre-training and fine-tuning module, an optimization module and a rolling optimization prediction module;

[0042] The data set construction module is used to simulate and generate the corresponding substation unit area building load data under different meteorological conditions and heat dissipation conditions based on the physical model of the target substation, and construct a data set;

[0043] The pre-training module is used to pre-train multiple load forecasting models based on the data set;

[0044] The optimization module is used to perform load forecasting based on multiple pre-trained load forecasting models, evaluate the load forecasting results, optimize the forecasting model based on the evaluation results, and obtain the final load forecasting model.

[0045] The rolling optimization prediction module is used to perform rolling optimization prediction on the final load prediction model based on the actual real-time data of the target substation.

[0046] An electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor is used to call and run the computer program stored in the memory to execute any of the methods described above.

[0047] A computer-readable storage medium stores a computer program, wherein the computer program implements the steps of any of the above methods when executed by a processor.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] The present invention proposes a substation load prediction method based on physical model pre-training. The method pre-trains the load prediction model through simulation data output by the physical model of the target substation, and then performs rolling optimization prediction on the load prediction model in combination with actual real-time data. In this way, the model can better adapt to the real environment while maintaining its adaptability to unknown data, thereby improving the generalization ability of the model; simulation data and actual data can provide different gradient information. By combining the gradient information of the two, not only can the parameters of the model be adjusted more effectively and the performance of the model be improved, but the model can also learn more load characteristics and laws, thereby improving the accuracy of the prediction.

[0050] The substation load forecasting method of the present invention evaluates the load forecasting results, and then optimizes the forecasting model according to the evaluation results, thereby obtaining the final load forecasting model; this design helps to identify which models are more suitable for load forecasting, thereby selecting the optimal model to improve the forecasting accuracy, and also helps to quickly determine which parameters have a greater impact on the model performance, so as to adjust the parameters in a targeted manner, thereby improving the calculation efficiency; BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flow chart of a substation load prediction method based on physical model pre-training in an embodiment;

[0052] Figure 2 This is a block diagram of the Transform load prediction principle in the embodiment;

[0053] Figure 3 This is a diagram of the LSTM neural network prediction architecture in the embodiment;

[0054] Figure 4 It is the RBF neural network prediction principle diagram in the embodiment;

[0055] Figure 5 Schematic diagram of support vector in the embodiment. DETAILED DESCRIPTION

[0056] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent modifications of the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0057] Embodiment 1:

[0058] A substation load forecasting method based on physical model pre-training, such as Figure 1 As shown, the specific steps are as follows:

[0059] S1. Based on the physical model of the target substation, simulate and generate the corresponding substation unit area building load data under different meteorological condition data and heat dissipation condition data, and construct a data set based on the meteorological condition data and the corresponding substation unit area building load data;

[0060] S2, pre-training multiple load forecasting models based on the data set;

[0061] S3. Perform load forecasting based on multiple pre-trained load forecasting models, evaluate the load forecasting results, optimize the forecasting model based on the evaluation results, and obtain the final load forecasting model;

[0062] S4. Perform rolling optimization forecasting on the final load forecasting model based on the actual real-time data of the target substation.

[0063] Embodiment 2:

[0064] This embodiment is further designed on the basis of the first embodiment in that the simulation of the corresponding substation unit area building load data under different meteorological data and heat dissipation data is performed using EnergyPlus energy consumption simulation software and Dest energy consumption simulation software.

[0065] Embodiment three:

[0066] This embodiment is further designed on the basis of the first embodiment in that the meteorological conditions in this example include temperature condition data, humidity condition data, air wet-bulb temperature condition data and wind speed condition data; the heat dissipation condition data is the heat generated by the equipment per unit time; the substation unit area building load data includes the electric energy consumption per unit area (EEC) and the air conditioning energy consumption (AEC).

[0067] Embodiment 4:

[0068] This embodiment is further designed on the basis of the third embodiment in that the specific method for pre-training multiple load forecasting models based on the data set in this embodiment is to execute sub-step 1 and sub-step 2 on the multiple load forecasting models respectively; wherein,

[0069] Sub-step 1:

[0070] Data set division: Sort the substation unit area building load data corresponding to the simulated data of different meteorological conditions according to the time series correlation, select the first 90% of the continuous data as the training set for model training; the remaining 10% of the data is used as the test set to verify the model performance; further, the continuous segment data in the test set can be selected as the verification set for detailed analysis and evaluation of the prediction accuracy. The continuous segment data can be a total of 168 hours of segment data in a week. Data normalization: The data normalization step is implemented to eliminate the adverse effects of abnormal data on model training. This process maps the range of sample data to the [0,1] interval through a specific algorithm, improves the uniformity of data processing, and accelerates the convergence speed during neural network training. The specific calculation formula for normalization is as follows:

[0071]

[0072] Among them, X norm is the normalized value, Xmin and Xmax are the minimum and maximum values ​​of the data respectively;

[0073] Sub-step 2:

[0074] The load forecasting model is pre-trained using the meteorological condition data in the training set as the model input data and the substation unit area building load data as the model output data;

[0075] The test set is used to verify the performance of the pre-trained load forecasting model.

[0076] The following is an example to illustrate the data set. This example selects the substation building energy consumption simulation load data corresponding to different meteorological conditions from the 1st hour to the 7884th hour of the whole year. Some of the data are shown in the following table:

[0077]

[0078]

[0079] Embodiment five:

[0080] This embodiment is further designed on the basis of the first embodiment in that the load forecasting model in this embodiment includes a Transformer model, an LSTM neural network, a support vector machine, and an RBF neural network; wherein the Transformer model is based on a seq2seq architecture, such as Figure 2As shown in the figure, it consists of two parts: an encoder and a decoder. The encoder converts the input sequence into a series of continuous vector representations (also called context vectors), and the decoder generates the target sequence based on the vector representation.

[0081] Specifically, the encoder is composed of multiple encoder layers, each of which contains a self-attention mechanism and a feedforward neural network. The self-attention mechanism allows the model to consider other elements in the sequence when processing each element in the sequence.

[0082] Decoder: It is also composed of multiple decoder layers stacked together. In addition to the self-attention mechanism and feedforward neural network, each decoder layer also contains an encoder-decoder attention mechanism to combine the output of the encoder with the input of the decoder.

[0083] Output each element in the sequence (the forecast load at time t) is regarded as the model's response to the input sequence X = x 1 ,x 2 ,…,x t-1 ],Right now

[0084] Residual connections and layer normalization: In order to improve the training efficiency and stability of the model, the Transformer model introduces residual connections and layer normalization in the encoder and decoder layers.

[0085] The LSTM neural network consists of multiple stacked LSTM layers and one or more fully connected layers (also called dense layers). An appropriate number of units is set in each LSTM layer, that is, the size of the hidden layer. The LSTM layer can handle dependencies in sequence data; stacking LSTM layers is to increase the complexity and learning ability of the model.

[0086] LSTM uses input gate, forget gate and output gate. These gated units control the flow of information through learning. The specific prediction architecture is as follows Figure 3 shown.

[0087] (1) Calculate the information discarded from the cell state at the previous moment by the forget gate. Based on the hidden state at the previous moment (h t-1 ) and the current input (x t ) to make a decision.

[0088] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0089] Where: ft is the output of the forget gate, indicating the degree of forgetting of each cell state value (0 means complete forgetting, 1 means complete retention). σ is the Sigmoid function, which maps any value to the (0,1) interval; W f and b f are the weight and bias of the forget gate, which are the parameters learned during training; [h t-1 ,x t ] means concatenating the hidden state of the previous moment and the input of the current moment.

[0090] (2) The calculation input gate determines which new information will be stored in the cell state, which includes two parts: one part is the output of the input gate, which determines which values ​​will be updated; the other part is the candidate cell state, which contains the new candidate values.

[0091] Output of the input gate:

[0092] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0093] Candidate cell states:

[0094]

[0095] Where: i t is the output of the input gate, which determines which values ​​will be updated; is the candidate cell state, including the new candidate value; W i 、b i , W C and b C are the weights and biases of the input gate and candidate cell states.

[0096] (3) The computational output gate is responsible for deciding which parts of the cell state will be output to the hidden state, based on the cell state and the input at the current moment.

[0097] o t =σ(W o ·[h t-1 ,x t ]+b 0 )

[0098] h t =o t *tanh(C t )

[0099] Among them: t is the output of the output gate, which determines which parts of the cell state will be output; h tis the hidden state at the current moment and the output of the LSTM unit; W o and b o are the weight and bias of the output gate.

[0100] The RBF neural network consists of an input layer, a hidden layer (also called a radial basis layer), and an output layer. The hidden layer neurons use radial basis functions as activation functions; Figure 4 As shown in the figure, the input layer receives input data (such as the building load data per unit area of ​​the substation under different meteorological conditions and heat dissipation conditions); each hidden layer neuron uses the radial basis function to calculate the distance between the input data and the center point, and generates output accordingly. The formula of the Gaussian function as the radial basis function can be expressed as:

[0101]

[0102] Among them, x is the input data, c i is the center point of the i-th hidden layer neuron, σ i is the width parameter of the neuron, and ∥·∥ represents the Euclidean distance.

[0103] In the output layer, the output of the hidden layer is linearly combined to obtain the final prediction result. The calculation formula of the output layer can be expressed as:

[0104]

[0105] Among them, y is the prediction result, N is the number of neurons in the hidden layer, and w i is the weight from the i-th hidden layer neuron to the output layer, and b is the bias term;

[0106] During the training phase, the prediction error is minimized by adjusting the RBF neural network parameters (including the center point ci, width σi, weight wi, and bias b).

[0107] In load forecasting, the load value is regarded as a continuous target variable, and a regression version of SVM, Support Vector Regression (SVR), is used. SVR is represented by a function f(x) that can fit the training data as closely as possible while maintaining a certain generalization ability. SVR achieves this by introducing an ε-insensitive loss function, that is, for each training point (xi, yi), if the absolute value of the difference between the predicted value f(xi) and the true value yi is less than ε, then the point is considered to have no loss.

[0108] like Figure 5 As shown, the optimization problem of support vector machine is expressed as:

[0109]

[0110] Among them, w is the weight vector, b is the bias term, C is the regularization parameter used to control the balance between model complexity and training error, and ξi and ξi* are slack variables used to deal with noise and outliers in the data.

[0111] By solving the optimization problem, we get the weight vector w and the bias term b, and then get the prediction function f(x) = w x + b. However, in practical applications, in order to deal with nonlinear problems and improve the generalization ability of the model, kernel techniques are usually used to map the input space to a high-dimensional feature space and find the optimal hyperplane in the high-dimensional feature space. At this time, the prediction function becomes:

[0112]

[0113] Among them, K(x,xi) is the kernel function used to calculate the similarity between x and xi; αi and αi* are Lagrange multipliers, which correspond to the support vectors in the training set.

[0114] Embodiment six:

[0115] This embodiment is further designed on the basis of the first embodiment in that the specific steps of evaluating the load forecasting result in this embodiment and optimizing the forecasting model based on the evaluation result to obtain the final load forecasting model include:

[0116] The mean square error (MSE), mean absolute error (MAE) and R2 score (also known as the coefficient of determination) are used to evaluate the load forecasting results of different load forecasting models, and then the load forecasting model with the best evaluation result is selected as the final load forecasting model based on actual needs, or multiple load forecasting models are combined for prediction and weights are assigned according to the load forecasting results.

[0117] Among them, the expression of mean square error (MSE) is:

[0118]

[0119] Among them: yi is the true value, is the predicted value and n is the number of samples.

[0120] The expression of mean absolute error (MAE) is:

[0121]

[0122] The expression for the R2 score (also called the coefficient of determination) is:

[0123]

[0124] in: is the average of the true values.

[0125] When combining predictions, the prediction results of multiple prediction methods are weighted averaged according to certain weights. Let the prediction value of the jth prediction method be The weight is ω j , then the formula for combined prediction is:

[0126]

[0127] in: represents the result of combined prediction; m represents the number of prediction models involved in the combination; ω j Represents the weight of the j-th prediction model, satisfying That is, the sum of the weights of all models is 1; Represents the prediction result of the jth prediction model. The weight can be determined in the following way:

[0128] Error-based weight allocation: weights are allocated based on the historical prediction errors of each model. Models with smaller errors are assigned larger weights, and models with larger errors are assigned smaller weights. After allocation, historical data is used to evaluate the errors of each model.

[0129] Optimization algorithm solution: Take the weight as the optimization variable and solve the optimal weight by minimizing the error of the combined prediction (such as MSE, RMSE, etc.).

[0130] Embodiment seven:

[0131] This embodiment is further designed on the basis of the first embodiment in that the specific method for performing rolling optimization prediction on the final load prediction model based on the actual real-time data of the target substation in this embodiment includes the following sub-steps 1 and 2; wherein,

[0132] Sub-step 1, using rolling forecasts for data validation, includes:

[0133] Rolling forecast is a dynamic data verification method suitable for time series data prediction problems. After each prediction, new observation data (that is, actual data) is added to the training set, and then the model is retrained based on the updated training set and the next prediction is made. In this way, rolling forecast can evaluate the generalization ability and adaptability of the model when processing new data.

[0134] The prediction formula used in the rolling forecast represents the behavior of the model in each forecast. The prediction formula adopts the moving average formula or the exponential smoothing formula. The moving average formula is as follows:

[0135]

[0136] Among them, x iIt is historical data and the number of periods of moving average.

[0137] The formula for exponential smoothing is as follows:

[0138] S t =αx t +(1-α)S t-1

[0139] Predicted value t+1 =S t

[0140] Where St is the smoothed value of the tth period, xt is the actual value of the tth period, and α is the smoothing coefficient (0<α<1);

[0141] The moving average method is more suitable for short-term load forecasting, while the exponential smoothing method is more suitable for long-term load forecasting. When testing multiple models, select the appropriate method for rolling forecasting according to the specific situation;

[0142] Sub-step 2, using k-fold cross-validation to evaluate model performance, includes:

[0143] K-fold cross evaluation reduces the risk of overfitting and gives a reliable estimate of model performance by splitting the dataset multiple times and repeating the training and testing process.

[0144] Using k-fold cross validation, the average performance metric of the model (such as average RMSE) is calculated by the following formula:

[0145]

[0146] Where k is the number of subsets, n i is the number of samples in the i-th test set, Y ij is the actual value of the jth sample in the i-th test set, is the corresponding predicted value. This formula calculates the average RMSE on all test sets as the final evaluation indicator of model performance.

[0147] Embodiment eight:

[0148] A substation load forecasting system based on physical model pre-training, including a data set building module, a pre-training and fine-tuning module, an optimization module and a rolling optimization prediction module;

[0149] The data set construction module is used to simulate and generate the corresponding substation unit area building load data under different meteorological conditions and heat dissipation conditions based on the physical model of the target substation, and construct a data set;

[0150] A pre-training module, used to pre-train multiple load forecasting models based on data sets;

[0151] The optimization module is used to perform load forecasting based on multiple pre-trained load forecasting models, evaluate the load forecasting results, optimize the forecasting model based on the evaluation results, and obtain the final load forecasting model.

[0152] The rolling optimization prediction module is used to perform rolling optimization prediction on the final load prediction model based on the actual real-time data of the target substation.

[0153] Embodiment nine:

[0154] An electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor is used to call and run the computer program stored in the memory to execute any of the methods described above.

[0155] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of any of the above methods are implemented.

[0156] The above description is only a specific implementation of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a technician familiar with the technical field within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. A substation load forecasting method based on physical model pre-training, characterized in that: include: Based on the physical model of the target substation, simulate and generate the corresponding substation unit area building load data under different meteorological conditions and heat dissipation conditions, and construct a data set based on the meteorological conditions and the corresponding substation unit area building load data; Pre-training multiple load forecasting models based on the data set; Perform load forecasting based on multiple pre-trained load forecasting models, evaluate the load forecasting results, optimize the forecasting model based on the evaluation results, and obtain the final load forecasting model; The final load forecasting model is subjected to rolling optimization forecasting based on the actual real-time data of the target substation.

2. The substation load prediction method based on physical model pre-training according to claim 1 is characterized in that: The simulation to generate the corresponding substation unit area building load data under different meteorological condition data and heat dissipation condition data is performed using EnergyPlus energy consumption simulation software and Dest energy consumption simulation software.

3. The substation load prediction method based on physical model pre-training according to claim 1 is characterized in that: The meteorological condition data includes temperature condition data, humidity condition data, air wet bulb temperature condition data and wind speed condition data; the heat dissipation condition data is the heat generated by the equipment per unit time; the substation unit area building load data includes the electricity consumption per unit area and the air conditioning energy consumption.

4. The substation load prediction method based on physical model pre-training according to claim 3 is characterized in that: The specific method for pre-training multiple load forecasting models based on the data set is to perform sub-step 1 and sub-step 2 on multiple load forecasting models respectively; wherein, Sub-step 1: Dataset division: Sort the substation unit area building load data corresponding to the simulated meteorological conditions according to time series correlation, select the first 90% of the continuous data as the training set for model training; the remaining 10% of the data is used as the test set to verify the model performance; Data normalization: The specific calculation formula for normalization is as follows: Among them, X norm is the normalized value, Xmin and Xmax are the minimum and maximum values ​​of the data respectively; Sub-step 2: The load forecasting model is pre-trained using the meteorological condition data in the training set as the model input data and the substation unit area building load data as the model output data; The test set is used to verify the performance of the pre-trained load forecasting model.

5. The substation load forecasting method based on physical model pre-training according to claim 1 is characterized in that: The load forecasting model includes a Transformer model, an LSTM neural network, a support vector machine, and an RBF neural network; wherein the Transformer model is based on a seq2seq architecture and consists of an encoder and a decoder. The encoder converts an input sequence into a series of continuous vector representations, and the decoder generates a target sequence according to the vector representations. The LSTM neural network is composed of a plurality of stacked LSTM layers and one or more fully connected layers, and the number of units in each LSTM layer is set; The RBF neural network consists of an input layer, a hidden layer and an output layer. The hidden layer neurons use the radial basis function as an activation function. The input layer receives input data. Each hidden layer neuron uses the radial basis function to calculate the distance between the input data and the center point, and generates an output accordingly. The formula of the Gaussian function as the radial basis function can be expressed as: Among them, x is the input data, c i is the center point of the i-th hidden layer neuron, σ i is the width parameter of the neuron, ∥·∥ represents the Euclidean distance; In the output layer, the output of the hidden layer is linearly combined to obtain the final prediction result. The calculation formula of the output layer can be expressed as: Among them, y is the prediction result, N is the number of neurons in the hidden layer, and w i is the weight from the i-th hidden layer neuron to the output layer, and b is the bias term; The optimization problem of the support vector machine is expressed as: Among them, w is the weight vector, b is the bias term, C is the regularization parameter used to control the balance between the complexity of the model and the training error, ξi and ξi* are both slack variables used to deal with noise and outliers in the data; By solving the optimization problem, the weight vector w and the bias term b are obtained, and then the prediction function f(x)=w·x+b is obtained.

6. The substation load forecasting method based on physical model pre-training according to claim 1 is characterized in that: The specific steps of evaluating the load forecasting results, optimizing the forecasting model based on the evaluation results, and obtaining the final load forecasting model include: The mean square error, mean absolute error and R2 score are used to evaluate the load forecasting results of different load forecasting models, and then the model with the best evaluation result is selected as the final load forecasting model according to actual needs, or multiple load forecasting models are combined for prediction and weights are assigned according to the load forecasting results.

7. The substation load forecasting method based on physical model pre-training according to claim 1 is characterized in that: The specific method for performing rolling optimization prediction on the final load prediction model based on the actual real-time data of the target substation includes the following sub-steps 1 and 2; wherein: Sub-step 1: Use a forecast formula in the rolling forecast to represent the behavior of the model in each forecast. The forecast formula uses a moving average formula or an exponential smoothing formula. Sub-step 2: Using k-fold cross validation, the average performance index of the model is calculated by the following formula: Where k is the number of subsets, n i is the number of samples in the i-th test set, Y ij is the actual value of the jth sample in the i-th test set, is the corresponding predicted value.

8. A substation load forecasting system based on physical model pre-training, characterized in that: It includes dataset construction module, pre-training and fine-tuning module, optimization module and rolling optimization prediction module; The data set construction module is used to simulate and generate the corresponding substation unit area building load data under different meteorological conditions and heat dissipation conditions based on the physical model of the target substation, and construct a data set; The pre-training module is used to pre-train multiple load forecasting models based on the data set; The optimization module is used to perform load forecasting based on multiple pre-trained load forecasting models, evaluate the load forecasting results, optimize the forecasting model based on the evaluation results, and obtain the final load forecasting model. The rolling optimization prediction module is used to perform rolling optimization prediction on the final load prediction model based on the actual real-time data of the target substation.

9. An electronic device, characterized in that: The electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor is configured to call and run the computer program stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method as claimed in any one of claims 1 to 7 are implemented.

Images