Electric power flow prediction method based on neural network and boosting
By using neural networks and boosting methods, the power network flow sequence is decomposed, reconstructed, and adjusted. Combined with GRU-Xgboost and SVR models, the nonlinearity and time correlation of time series data in power network flow forecasting are solved, and higher accuracy flow forecasting and flow management strategy formulation are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF POSTS & TELECOMM
- Filing Date
- 2023-10-19
- Publication Date
- 2026-06-26
AI Technical Summary
Existing power load forecasting models cannot effectively utilize the temporal correlation and nonlinearity of time series data when dealing with large-scale and complex network traffic data, resulting in limited forecast accuracy, especially when facing the variability and abrupt changes in power networks.
A neural network-based boosting approach is used to decompose, reconstruct, and adjust the power network flow sequence. The GRU-Xgboost model is combined for feature reconstruction and time series prediction, and the residual sequence is processed by the SVR model to construct a power dispatch network flow prediction model, taking into account nonlinear characteristics and the time correlation of time series data.
It improves the accuracy and completeness of power grid traffic prediction, enabling more accurate prediction of network traffic change trends, meeting users' needs for network service quality, and enhancing the model's training speed and prediction accuracy.
Smart Images
Figure CN117424217B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of power load forecasting technology and network traffic forecasting, specifically relating to a power traffic forecasting method based on neural networks and boosting. Background Technology
[0002] In the 21st century, digital technologies have become deeply integrated with power companies, leading to a continuous growth in digital business lines, a dramatic increase in the scale and operational complexity of IT systems, and an explosive growth in network traffic. Effective network management has become crucial to prevent network congestion, improve resource utilization, and meet user demands for Quality of Service (QoS). By modeling and predicting network traffic, we can anticipate traffic trends and formulate appropriate traffic management strategies.
[0003] The power dispatch data network is the foundation of power grid dispatch automation and modern management, ensuring the safe, stable, and economical operation of the power grid. It plays a crucial role in coordinating various components of the power system, such as generation, transmission, substation, distribution, and consumption, ensuring the safe, economical, stable, and reliable operation of the power grid. Therefore, processing and predicting large-scale, complex network traffic data is essential in the power dispatch network. Currently, to address the demand for processing massive time-series data, short-term load flow forecasting based on big data is widely used in power dispatch. However, due to the nonlinear characteristics of modern network traffic, such as its abrupt changes and multiple dependencies, traditional linear models are suitable for small-scale, sparse network traffic but cannot meet the requirements. Prediction models based on machine learning algorithms can better explain the randomness and periodicity of network traffic, such as those described in patent applications CN112766078A and CN109978217A, but they typically ignore the temporal correlation of time-series data, resulting in limited prediction accuracy. Summary of the Invention
[0004] To address the aforementioned technical problems, this invention provides a power flow forecasting method based on neural networks and boosting. It decomposes, reconstructs, and adjusts complex network flow time series, proposing a novel network flow time series forecasting model that considers nonlinear characteristics and the temporal correlation of time series data, thereby improving forecast accuracy. Considering the variability of power networks and based on the operational characteristics of electricity, it identifies the characteristic factors affecting the flow model and effectively handles the nonlinear complexity and unknown noise of the flow series, thus establishing a flow forecasting model for power networks. This model aims to provide forecasting information for relevant departments to make informed decisions.
[0005] The power flow prediction method based on neural networks and boosting described in this invention includes the following steps:
[0006] S1. Select two highly correlated power network traffic sequences, one of which is complete and the other is missing. Reconstruct the missing values of the power network traffic sequences using transfer learning.
[0007] S2. Construct a GRU-Xgboost model to reconstruct features from the power network flow sequence after missing value reconstruction, and perform time series prediction on the reconstructed features; combine the predicted flow data sequence output by the GRU-Xgboost model after training, and aggregate the predictions of all sub-sequences to obtain the power dispatch network flow prediction sequence.
[0008] S3. Difference the aggregated predicted values and time observations to obtain the residual sequence; construct the SVR residual sequence model, and generate the power dispatch network flow prediction compensation value based on the residual series. Simultaneously calculate the remaining series R; for Series and Summing these values yields the final power dispatch network flow forecast.
[0009] Furthermore, step 1 specifically involves:
[0010] Step 1-1: Train a basic GRU model using the rolling window method on the power dispatch network flow sequence with complete data. This basic model is used to process the source domain model with complete data.
[0011] Steps 1-2: Freeze the first 6 layers of the basic GRU model to ensure that the learned features are not destroyed; use the target power dispatch network flow sequence training samples with missing values DT for transfer learning.
[0012] Steps 1-3: Fine-tune the model parameters using the validation example in the target power dispatch network flow sequence DT to adapt to the data characteristics of the target domain;
[0013] Steps 1-4: Estimate the transmission input based on the data in the target power dispatch network flow sequence and the basic power dispatch network flow sequence;
[0014]
[0015]
[0016] Where r is the size of the scroll window. and These represent the data in the target power dispatch network flow sequence and the basic power dispatch network flow sequence, respectively.
[0017] The transmission input value is adjusted based on the average value of the target power dispatch network flow sequence and the basic power dispatch network flow sequence, as shown in the following formula:
[0018]
[0019] in The adjusted data of the basic power dispatch network flow sequence, This represents the average value of the flow sequence in the basic power dispatching network. M represents the average value of the target power dispatch network flow sequence. T This is a prediction model.
[0020] Steps 1-5: Repeat steps 1-1 to 1-4, reversing the time series and traversing backward from the most recent time point. At each time step, use the known positive estimation results to fill the missing values at that time point, and then continue to move forward one time step, using the positive estimation results of the previous time point to fill the missing values of the previous time point, and so on, until the beginning of the time series is reached.
[0021] Furthermore, based on GRU in the GRU-Xgboost model, feature reconstruction is performed on the power network flow sequence after reconstruction of missing values, specifically as follows:
[0022] Given a power dispatch network flow sequence containing N samples, defined as:
[0023] {(x i ,y i (4) | i = 1, 2, 3, ..., N}
[0024] For target y j Construct the input feature vector F j ,Right now:
[0025] F j ={x j ,x j+1 ,…,x j+m},j=1,2,3,…,Nm (5)
[0026] Where m is the value that will affect the x-th node. j+m Parameters for each data item;
[0027] The corresponding power dispatch network flow prediction output y j The following is given:
[0028] y j ={x j+m},j=1,2,3,…,Nm (6)
[0029] After the original sequence is input into the GRU model, the input vector and label are first defined. After dividing the training set and the test set, for each set of input vector and label, GRU obtains an improved feature vector, which is the hidden state in GRU, thereby completing the feature reconstruction of the original power network traffic sequence.
[0030] The specific update formula for GRU is as follows:
[0031] z=σ(W z x+U z h t-1 (7)
[0032] r=σ(W r x+U z h t-1 (8)
[0033]
[0034]
[0035] Where x represents F j Samples of power dispatch network flow data, h t-1 and h t Let W represent the previous and current hidden states, respectively, derived from the learned weight matrix. z U z W r U r W is generated by weighted average, and σ is the logical sigmoid function.
[0036] Furthermore, time series prediction is performed on the reconstructed features based on XGBoost in the GRU-Xgboost model, specifically as follows:
[0037] Suppose the tree model to be trained in the t-th iteration is f t (x), then we have:
[0038]
[0039] Where k is the number of base models;
[0040] The loss function can be derived from the flow forecast values of the power dispatching network. Compared with the true value Indicate:
[0041]
[0042] Where n represents the sample size of power network traffic data;
[0043] The prediction accuracy of a model is determined by both its bias and variance. The loss function represents the model's bias. To minimize variance, a regularization term needs to be added to the objective function to prevent overfitting. Therefore, the objective function consists of the model's loss function and a regularization term to suppress model complexity. The objective function is defined as follows:
[0044]
[0045] The second part of the above formula sums up the complexity of all t trees and adds it to the objective function as a regularization term to prevent the model from overfitting.
[0046] The model applies to the i-th power network flow data sample x i The predicted value is:
[0047]
[0048] In the above equation, the first part is the predicted value given in step t-1, and the second part is the predicted value of the new model to be added in step t. At this point, the objective function can be equivalently written as:
[0049]
[0050] Based on the formula transformation, the final target power dispatch network flow prediction function is transformed into a relationship between a variable and a constant.
[0051] The relationship between the second and third equal signs:
[0052] By splitting the regularization term, since the structure of the first tree is already determined, the sum of the complexities of the first tree can be represented by a constant, as shown below:
[0053]
[0054] Furthermore, in step 3, an SVR residual sequence model is constructed to obtain the optimal power dispatch network flow prediction, specifically as follows:
[0055] Given a training set of power dispatch network flow sequences containing N samples, a residual series R can be obtained that summarizes the power network flow predictions and actual observations:
[0056] R = y train -h train (12);
[0057] R = y train -h train (17)
[0058] Where y trainThese are the labels for the training set, where h represents the GRU-xgboost pair with y. train Operation;
[0059] These residuals are modeled and trained using Support Vector Regression (SVR) to adjust each aggregated power network flow forecast; the SVR model generates an adjusted forecast sequence. The two predictions are added together to obtain the final power network traffic prediction; the final prediction result is an aggregation of the GRU-xgboost and SVR prediction results.
[0060]
[0061] Where y represents the final prediction result. This represents the prediction results of GRU-xgboost. This represents the compensation value obtained from SVR.
[0062] The beneficial effects of this invention are as follows:
[0063] 1) This invention, through network traffic modeling and prediction, can grasp the changing characteristics and trends of network traffic in advance, thereby formulating reasonable and effective traffic management strategies to meet users' requirements for network service quality (QoS). It breaks through the limitations of traditional prediction models that can only predict small-scale sparse network traffic and achieves excellent results.
[0064] 2) This invention utilizes missing value reconstruction and model-based transfer learning to achieve higher initial model performance; during training, the model improves at a faster rate; after training, the model converges better; and it also improves the integrity of power network traffic data.
[0065] 3) The GRU-Xgboost model constructed in this invention can improve the quality of input feature vectors and generate more accurate network flow predictions, thereby improving prediction accuracy;
[0066] 4) This invention addresses the differences in power networks by identifying the characteristic factors affecting the flow model based on the business characteristics of electricity, and effectively handling the nonlinear complex characteristics and unknown noise of the flow sequence, thereby establishing a flow prediction model for power networks, aiming to provide predictive information for relevant departments to make decisions.
[0067] 5) The residual adjustment module based on the SVR model designed in this invention receives the prediction results of the GRU-XGBoost model and the actual observed power network flow data labels, performs residual calculation, generates a residual sequence (R) by calculating the difference between the actual labels and the GRU-XGBoost prediction values, and uses the support vector regression (SVR) model to model and train the residuals to further understand and adjust the model error, thereby improving the accuracy and reliability of power load data prediction. Attached Figure Description
[0068] Figure 1 This is a flowchart illustrating the method described in this invention;
[0069] Figure 2 This is a relatively complete power network traffic sequence obtained by applying the model-based transfer learning missing value reconstruction method in the example of this invention;
[0070] Figure 3 This is a power network traffic data prediction sequence generated by applying the GRU-XGboost model after reconstruction of missing values in an example of the present invention.
[0071] Figure 4 This is the final power network traffic data prediction sequence generated by the SVR residual adjustment module in this embodiment of the invention. Detailed Implementation
[0072] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.
[0073] See Figure 1 The power flow prediction method based on neural networks and boosting described in this invention includes the following steps:
[0074] S1, Missing Value Reconstruction: First, perform a transfer input. After identifying missing values in two highly correlated power network flow sequences (one complete, one missing), calculate the difference between their means, and subtract the difference between the means of the two sequences from the current value to obtain the transfer input to the result.
[0075] S2, then transfer learning is performed. The model trained using the complete power network traffic sequence is used to predict the missing value sequence, obtaining the final prediction result; finally, the final imputation result is shown; the training set part contains the results of the transfer input, and the test set part contains the prediction results of the transfer learning; such as... Figure 2 As shown, the lower part is the original sequence with missing values, and the upper part is the imputed sequence obtained after transfer learning. The missing value part is well improved, and the final prediction result (the imputed sequence) is close to the true value, indicating that the model has a good transfer effect.
[0076] S3, Research on Feature Reconstruction and Complex Time Series Prediction Algorithms Based on GRU and XGBoost: For power dispatch network traffic sequences reconstructed with missing values, GRU-XGBoost blocks are used to improve the input quality feature vector and obtain an improved reconstructed hidden state h. t For processed power network traffic data sequences, modeling is required for prediction.
[0077] S4. Therefore, by establishing the GRU-XGboost module, the original power dispatch network traffic data is divided into training and testing sets for prediction. In the GRU-XGboost model, GRU performs feature reconstruction of the traffic sequence, and XGboost performs time series prediction on the reconstructed data.
[0078] S5. The sequence reconstructed from missing values is input into the GRU model. To ensure the continuity of the subsequent predicted sequence, the definition of the overall training data and label function is completed first, then the training set and test set are divided, and finally the GRU model is used for feature reconstruction.
[0079] S6. The reconstruction result obtained by GRU is input into XGBoost. Similarly, the training data and label function are defined first, then the training set and test set are divided, and finally the time series prediction is completed.
[0080] S7 combines the two to form the GRU-Xgboost module, which outputs the predicted flow data sequence after training the GRU-XGboost model, thus obtaining the flow prediction sequence for the power dispatching network.
[0081] The overall prediction results of the GRU-xgboost module are as follows: Figure 3 As shown in the figure, the upper part represents the actual values, and the lower part represents the predicted values. The prediction effect is feasible. The GRU-Xgboost module performs feature reconstruction on the power network traffic sequence after the missing values are reconstructed. This means that the model not only focuses on predicting the original data, but also predicts the processed features, providing more information.
[0082] Design and Research of Residual Adjustment Module for S8 and SVR Models: Generating Power Dispatch Network Flow Prediction Compensation Values Based on Residual Series Using Residual Adjustment Blocks. Simultaneously, the remaining series R is calculated. Power dispatch network flow forecasting is the forecast of all decomposed subsequences obtained through aggregation. For the residual terms generated by decomposition, we designed a residual adjustment module to compensate for the impact of random noise;
[0083] S9, by differencing the aggregated predicted values and actual observed values, a residual sequence can be obtained. Then, SVR is applied to model this residual sequence. In this part, an SVR residual adjustment block is designed to compensate for the influence of noise. To ensure the magnitude of the final prediction result, the test set portion of the SVR should be consistent with the prediction results of the GRU-XGBoost module. They are all the same size.
[0084] S10, therefore, after defining the training data and label function for the residual sequence R, when dividing the SVR dataset, the size of its test set has been determined by the prediction result size of the GRU-XGBoost module, while the training set is the remaining part of the sequence. The obtained power network flow residual sequence R is further predicted using SVR to obtain the power dispatch network flow prediction compensation value.
[0085] S11, Final prediction results for the GRU-XGBoost module Results of residual adjustment for SVR model Summing these values yields the final power dispatch network flow forecast y.
[0086] Final prediction results are as follows Figure 4 As shown in the figure, the lower part represents the actual value and the upper part represents the predicted value. It can be seen from the figure that residual adjustment improves the overall prediction accuracy. The curves of the actual value and the predicted value fit together well. It corrects the error and provides a more accurate prediction.
[0087] The above description is merely a preferred embodiment of the present invention and is not intended to further limit the present invention. All equivalent changes made based on the description and drawings of the present invention are within the protection scope of the present invention.
Claims
1. A power flow prediction method based on neural networks and boosting, characterized in that, Includes the following steps: Step S1: Select two highly correlated power network traffic sequences, one of which is complete and the other is missing. Reconstruct the missing values of the power network traffic sequences using a transfer learning method. Step S2: Construct the GRU-Xgboost model, reconstruct the features of the power network flow sequence after the missing values are reconstructed, and perform time series prediction on the reconstructed features; The predicted flow data sequence is generated and trained using a GRU-Xgboost model. The predictions of all sub-sequences are then aggregated to obtain the flow prediction sequence for the power dispatching network. ; Step S3: Difference the summarized predicted values and time observations to obtain the residual sequence; construct the SVR residual sequence model, and generate the power dispatch network flow prediction compensation value based on the residual series. Simultaneously calculate the remaining series R; for Series and Summing these values yields the final power dispatch network flow forecast. Step S1 is as follows: Step S1-1: Train the basic GRU model using the rolling window method on the power dispatch network flow sequence with complete data; Step S1-2: Freeze the first 6 layers of the basic GRU model and use the target power dispatch network flow sequence training samples with missing values DT for transfer learning. Steps S1-3: Fine-tune the model parameters using the validation example in the target power dispatch network flow sequence DT to adapt to the data characteristics of the target domain; Step S1-4: Estimate the transmission input based on the data in the target power dispatch network flow sequence and the basic power dispatch network flow sequence; (1) (2) Where r is the size of the scroll window. and These represent the data in the target power dispatch network flow sequence and the basic power dispatch network flow sequence, respectively. The transmission input value is adjusted based on the average value of the target power dispatch network flow sequence and the basic power dispatch network flow sequence, as shown in the following formula: (3)； in The adjusted data of the basic power dispatch network flow sequence, This represents the average value of the flow sequence in the basic power dispatching network. This represents the average value of the target power dispatch network flow sequence. For prediction models; Step S1-5: Repeat steps S1-1 to S1-4 to reverse the time series and traverse backward from the most recent time point. At each time step, use the known positive estimation results to fill the missing values at that time point, and then continue to move forward one time step, using the positive estimation results of the previous time point to fill the missing values of the previous time point, and so on, until the beginning of the time series is reached.
2. The power flow prediction method based on neural networks and boosting according to claim 1, characterized in that, Based on the GRU in the GRU-Xgboost model, feature reconstruction is performed on the power network flow sequence after reconstruction of missing values, specifically as follows: Given a power dispatch network flow sequence containing N samples, defined as: ; (4) For the goal Constructing input feature vectors ,Right now: (5) in This will affect the first Parameters for each data item; Corresponding power dispatching network flow forecast output The following is given: (6)； After the original sequence is input into the GRU model, the input vector and label are first defined. After dividing the training set and the test set, for each set of input vector and label, GRU obtains an improved feature vector, which is the hidden state in GRU, thereby completing the feature reconstruction of the original power network traffic sequence. The specific update formula for GRU is as follows: (7) (8) (9) (10) in, express Samples of power dispatch network flow data in China and Let represent the previous and current hidden states, respectively, derived from the learned weight matrix. , , , , Weighted generation, It is the logical sigmoid function.
3. The power flow prediction method based on neural networks and boosting according to claim 1, characterized in that, The reconstructed features are used for time series prediction based on XGBoost in the GRU-Xgboost model, specifically as follows: Assume the first The tree model to be trained in the next iteration is Then we have: (11) in, It is the number of base models; The loss function is based on the flow prediction values from the power dispatching network. Compared with the true value Indicate: (12) in, This indicates the sample size of power network traffic data; The objective function is defined as follows: (13) The second part of the above formula sums up the complexity of all t trees and adds it to the objective function as a regularization term. For the first A sample of power network traffic data The predicted value is: (14) In the above equation, the first part is the predicted value given in step t-1, and the second part is the predicted value of the new model that needs to be added in the first step. At this time, the objective function is equivalently written as: (15) Based on the formula transformation, the final target power dispatch network flow prediction function is transformed into a relationship between a variable and a constant. The relationship between the second and third equal signs: The regularization term is split into its components. Since the structure of the first tree is already determined, the sum of the complexities of the first tree can be represented by a constant, as shown below: (16)。 4. The power flow prediction method based on neural networks and boosting according to claim 1, characterized in that, In step S3, an SVR residual sequence model is constructed to obtain the optimal power dispatch network flow prediction, specifically as follows: Given a training set of power dispatch network traffic sequences containing N samples, a residual series R is obtained that summarizes the power network traffic predictions and actual observations: (17) in These are the labels for the training set. This indicates that GRU-xgboost supports... Operation; These residuals are modeled and trained using Support Vector Regression (SVR) to adjust each aggregated power network flow forecast; the SVR model generates an adjusted forecast sequence. The two predictions are added together to obtain the final power network traffic prediction; the final prediction result is an aggregation of the GRU-xgboost and SVR prediction results: (18) Where y represents the final prediction result. This represents the prediction results of GRU-xgboost. This represents the compensation value obtained from SVR.