Load prediction method and system based on diversity regularization stacking learning, medium and product

Abstract

The invention discloses a load prediction method and system based on diversity regularization stacking learning, a medium and a product, and relates to the technical field of power load prediction, and the method comprises the steps: obtaining power operation historical data for training; performing correlation screening on the historical power operation data for training by using an embedded feature selection method and a correlation analysis method to obtain screened historical power operation data; performing diversity regularization pruning on the sub-model pool to obtain a target basic learner combination; the sub-model pool comprises a plurality of basic learners for load prediction; carrying out stacking training on the target basic learner combination according to the screened power operation historical data to obtain a composite predictor; and performing load prediction on the to-be-predicted power operation historical data by using the composite predictor to obtain a load prediction result. The generalization ability of the load prediction model is enhanced, and the accuracy of power system load prediction is improved.

Description

Technical Field

[0001] The present application relates to the technical field of electric load forecasting, and particularly to a load forecasting method, system, medium and product based on diversity regularization stacking learning. Background Art

[0002] The electricity market is a management mechanism and execution system that uses economic, legal and other means to organize and coordinate the operation of members in various links of the power system, such as power generation, transmission, distribution and power consumption, in accordance with the principles of fair competition, voluntariness and mutual benefit. The traditional electricity industry mainly adopts a vertically integrated monopoly operation model, where a single utility company is responsible for the entire process from power generation to distribution. The adjustment of energy policies globally has promoted the development of the electricity market. Many countries and regions have introduced energy market reform policies, such as the electricity market liberalization policy in the European Union. The aim is to break the traditional electricity monopoly and promote the access of renewable energy and the utilization of distributed energy resources. The rapid development of smart grid technology provides technical support for the electricity market. Smart grid can achieve two-way communication, real-time monitoring and control of the power system, improving the reliability and flexibility of the power grid.

[0003] Technologies based on artificial intelligence are being widely developed and deployed in real-world tasks. Due to their reasoning ability and interpretability, they have brought extensive revolutionary changes in the field of load forecasting. The energy industry has accumulated rich data assets, creating innovative opportunities for the application of data-driven technologies. However, limited by the application scope of a single algorithm, the artificial intelligence models used for load forecasting have the problem of weak generalization, further increasing the difficulty of load forecasting work. Load forecasting, as one of the fields where artificial intelligence is widely applied in electrical engineering, has been a research hotspot since the emergence of the power industry. In the past few decades, the quality of electric load forecasting has made a huge contribution to the efficiency of power companies, affecting decisions such as energy purchase plans, load dispatching, demand response and system risk assessment. The accurate and adaptive forecasting of demand-side load plays a key role in power system generation, planning, dispatching and marketing. Currently, there are numerous professional literatures on load forecasting, and the research mainly focuses on key issues such as forecasting scenarios, time ranges, feature selection, model training, hyperparameter and parameter optimization. In terms of load forecasting scenarios, load forecasting tasks can be divided into system load types, microgrid load types, electric vehicle station load types, etc. The load profiles in different forecasting tasks have different characteristics. Due to the cancellation effect among multiple individual users, the load distribution of a large-capacity power system is usually stable. In contrast, the load profile changes in the distribution network or microgrid are often more random, and the corresponding forecasting tasks are more difficult, and complex scenarios pose higher requirements for model construction.

[0004] At present, single models are mostly used for artificial intelligence load forecasting in large-capacity power systems, resulting in problems such as poor generalization ability of load forecasting models and inaccurate load forecasting of power systems. Summary of the Invention

[0005] The purpose of this application is to provide a load forecasting method, system, medium and product based on diversity regularization stacking learning, which can enhance the generalization ability of the load forecasting model and improve the accuracy of power system load forecasting.

[0006] To achieve the above purpose, the present application provides the following solutions.

[0007] In the first aspect, the present application provides a load forecasting method based on diversity regularization stacking learning. The load forecasting method based on diversity regularization stacking learning includes: obtaining historical power operation data for training; the historical power operation data specifically includes: historical load, weather data and calendar rules; using an embedded feature selection method and a correlation analysis method to perform relevance screening on the historical power operation data for training to obtain the screened historical power operation data; performing diversity regularization pruning on the sub-model pool to obtain a target base learner combination; the sub-model pool includes multiple base learners for load forecasting; performing stacking training on the target base learner combination according to the screened historical power operation data to obtain a composite predictor; using the composite predictor to perform load forecasting on the historical power operation data to be predicted to obtain a load forecasting result.

[0008] In the second aspect, the present application provides a computer system, including: a memory, a processor, and a computer program stored on the memory and executable on the processor. The processor executes the computer program to implement the load forecasting method based on diversity regularization stacking learning described in any one of the above.

[0009] In the third aspect, the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the load forecasting method based on diversity regularization stacking learning described in any one of the above.

[0010] In the fourth aspect, the present application provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the load forecasting method based on diversity regularization stacking learning described in any one of the above.

[0011] According to the specific embodiments provided by the present application, the following technical effects are disclosed in the present application.

[0012] This application improves the correlation of training data by using an embedded feature selection method and a correlation analysis method to screen the correlation of historical power operation data for training; obtains a target base learner combination by performing diversity regularization pruning on the sub-model pool, improving the generalization ability of the target base learner combination; obtains a composite predictor by performing stacking training on the target base learner combination, and uses the composite predictor to obtain a load prediction result, improving the accuracy of power system load prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0014] Figure 1 Schematic flowchart of a load prediction method based on diversity regularization stacking learning provided by an embodiment of the present application Figure 1 。

[0015] Figure 2 Schematic flowchart of a load prediction method based on diversity regularization stacking learning provided by an embodiment of the present application Figure 2 。

[0016] Figure 3 Performance comparison chart of the stacking model and the single model for ultra-short-term power load prediction provided by an embodiment of the present application.

[0017] Figure 4 Performance comparison chart of the stacking model and the single model for ultra-short-term power consumption prediction provided by an embodiment of the present application.

[0018] Figure 5 Schematic structural diagram of a computer system provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, rather than all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0020] To make the above objects, features, and advantages of the present application more obvious and understandable, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0021] In the process of artificial intelligence load forecasting, existing feature analysis algorithms have potential drawbacks such as being insensitive to non-linear correlations, having long execution times, and lacking model performance orientation, making it difficult to effectively quantify the importance of features. Moreover, most integrated load forecasting models work randomly and blindly in the selection of basic learners, mostly combining all of them or randomly combining them, and unable to systematically select useful subsets of individual learners. There is a lack of an integrated framework that can accommodate various sub-models in the field of load forecasting. Among them, homogeneous sub-models are difficult to observe data from different perspectives, and their generalization and prediction performance are poor. Existing research on ensemble learning only selects a few algorithms to build load forecasting models, and the discussion on diversity is still insufficient. Most mature load forecasting models are usually deployed by a single algorithm, and a single algorithm cannot reflect the true distribution of feature importance.

[0022] Embodiment 1, as Figure 1 - Figure 2 shown, this embodiment provides a load forecasting method based on diversity regularization stacking learning. The load forecasting method based on diversity regularization stacking learning includes the following steps.

[0023] S1. Obtain historical power operation data for training; the historical power operation data specifically includes: historical load, weather data, and calendar rules.

[0024] S2. Use the embedded feature selection method and the correlation analysis method to screen the historical power operation data for training to obtain the screened historical power operation data.

[0025] Further, step S2 specifically includes the following steps.

[0026] S21. According to the historical power operation data for training, use the embedded feature selection method to obtain the contribution degree of the target predicted load; the contribution degree of the target predicted load includes: the contribution degree of the historical load, the contribution degree of the weather data, and the contribution degree of the calendar rules.

[0027] Optionally, the embedded feature selection method includes the XGBoost tree boosting method, the LightGB tree boosting method, and the CatBoost tree boosting method.

[0028] In the actual application process, the process of step S21 is as follows: According to the historical power operation data for training, use multiple tree boosting methods such as XGBoost, LightGB, and CatBoost to obtain the contribution degree of the target predicted load in a predictive programming manner.

[0029] S22. Analyze the correlations between the target predicted load and historical load, weather data, and calendar rules respectively using the correlation analysis method based on the target predicted load contribution degree, and obtain the target predicted load correlation degree; the target predicted load correlation degree includes: the correlation degree of historical load, the correlation degree of weather data, and the correlation degree of calendar rules.

[0030] S23. Screen the historical power operation data for training according to the target predicted load correlation degree to obtain the screened historical power operation data.

[0031] S3. Perform diversity regularization pruning on the sub-model pool to obtain a target basic learner combination; the sub-model pool includes multiple basic learners for load prediction.

[0032] Further, step S3 specifically includes the following steps.

[0033] S31. Calculate the prediction generalization error of each basic learner in the sub-model pool.

[0034] Further, the calculation formula for the prediction generalization error of the basic learner is as follows.

[0035]

[0036] In the formula, e(l) is the generalization error of basic learner l, b 2 (l) is the bias of basic learner l, v(l) is the variance of basic learner l, μ 2 is the irreducible error, E(l) is the expected output of basic learner l, E(·) is the expectation operator, Q(l) is the predicted value of basic learner l, and f is the true function.

[0037] S32. Calculate and determine the individual accuracy of each basic learner based on the prediction errors of each basic learner.

[0038] S33. Use the mutual information theory and hierarchical clustering to perform regularization diversity analysis on the basic learner combination according to the individual accuracy of each basic learner to obtain the target basic learner combination.

[0039] Step S33 specifically includes the following steps.

[0040] S331. Analyze and obtain the mutual information coefficient using the mutual information theory according to the individual accuracy of each basic learner.

[0041] Optionally, the mutual information coefficient is 0.5.

[0042] S332. Search from the root of the dendrogram using the mutual information coefficient to determine the clusters most suitable for pruning the base learners and diversity regularization; the dendrogram is obtained by hierarchical clustering of the base learners according to the prediction errors; the starting node of the dendrogram is the root of the dendrogram, representing the whole of the base learners.

[0043] S333. Screen and combine the target base learner combination from multiple base learners according to the clusters most suitable for pruning the base learners and diversity regularization.

[0044] In the actual application process, the actual application process of step S3 is as follows.

[0045] First, according to the performance of the sub-models (base learners) in the prediction task, use the mutual information theory and hierarchical clustering to evaluate the degree of diversity of regularization. Use mutual information as a statistical indicator to assign scores to the error correlations of the sub-models. Calculate the error correlations through the scores of mutual information to identify heterogeneous sub-models, and filter out the base learners with the largest differences. The lower the mutual information score, the lower the correlation between the base learners. At the same time, use hierarchical clustering as an unsupervised clustering method to group the base learner errors according to similar patterns, and put them into clusters by creating a dendrogram.

[0046] Second, evaluate the prediction ability of the base learners by sorting the individual accuracies, and calculate the mutual information and hierarchical clustering through the prediction errors of the individual base learners.

[0047] Third, select the sub-model with the highest individual accuracy in a dendrogram cluster as the representative base learner. And set the mutual information coefficient less than 0.5 as the division threshold of hierarchical clustering, that is, the mutual information coefficient between the representative base learners cannot exceed 0.5.

[0048] Fourth, start searching from the root of the dendrogram for the result of mutual information calculation to determine the position where the hierarchical tree is cut into clusters most suitable for pruning the base learners and diversity regularization. So that the selected learners not only have the highest accuracy in the dendrogram cluster, but also have the smallest correlation with each other.

[0049] Fifth, obtain the target base learner combination (the best base learner combination) by gradually excluding the base learner combinations with the worst accuracy, and use the following metrics to evaluate the model performance.

[0050]

[0051] e X =max[c(k)-c′(k)].

[0052] In the formula, e Mis the mean absolute percentage error, e R is the root mean square error, and e X is the maximum error, m is the total number of samples, and c(k) and c′(k) represent the k-th actual value and predicted value, respectively.

[0053] S4. Stacked training is performed on the target base learner combination according to the filtered historical power operation data to obtain a composite predictor.

[0054] Step S4 specifically includes the following steps.

[0055] S41. The filtered historical power operation data is divided into multiple data sets by the leave-one-out method; each data set includes multiple first training sets and one test set.

[0056] Step S41 specifically includes: The original training data is split into multiple data blocks, and each time one different data block is selected as the test set, and the remaining data blocks are used as the first training set to obtain multiple data sets.

[0057] S42. Stacked training is performed on the target base learner combination using multiple first training sets in different data sets to obtain the trained target base learner.

[0058] S43. The test sets in different data sets are input into the trained target base learner to obtain multiple single-model prediction results.

[0059] S44. A second training set is constructed according to the multiple single-model prediction results.

[0060] S45. The meta-learner is trained using the second training set to obtain a composite predictor.

[0061] In the actual application process, the specific implementation process of step S4 is as follows.

[0062] First step, select one data block from the filtered historical power operation data as the test set (test data set), and the other (remaining) data blocks as the first training set (training data set), and the instances used to generate the new data set are excluded from the training examples of the single learner.

[0063] Second step, use non-overlapping sub-data blocks to train a group (multiple) of base learners in the first layer.

[0064] Specifically, for the data set T = {(r m , u m ), m = 1,..., m}, r m is the feature vector of the m-th sample. u mis the predicted value corresponding to the m-th sample. h is the number of features. Each feature vector can be described as (r1, r2,..., r n ), and T1, T2,..., T i are i data blocks split from T and the number is the same as h. T i and T -i (T -i = T - T i ) are defined as the i-th fold test set (test data set) and the first training set (training data set), respectively. For the first layer, the base learner L i can be trained with the i-th algorithm on the data set T -i .

[0065] In the third step, the load prediction results from the base learners are combined into a second training set (new data set) with the same size as the filtered historical power operation data.

[0066] For each sample r i in the i-fold test data set T m , the result predicted by the base learner L i can be denoted as p im after the cross-validation process is completed.

[0067] The output is constructed into a new data set, T new = {(u m , p 1m ,..., p im ), m = 1,..., m}. In the first-level stacking framework, the data blocks used to test each base learner in the model are avoided from participating in the training of the learner.

[0068] In the fourth step, the output of the first-layer base learners is used as the input to train the meta-learner to obtain the composite predictor.

[0069] The data set generated by the base learner is the input data of the meta-learner. The second-level meta-learner L new can be trained through the new data set T new = {(u m , p 1m ,..., p im ), m = 1,..., m new}. Therefore, all the results from individual learners are combined into a second training set (new data set) with the same size as the filtered historical power operation data, realizing the feature transformation from the base learner to the meta-learner. The output of the base learner can be fully used in the induction process of the meta-learner.

[0070] S5. Use a composite predictor to perform load forecasting on the historical power operation data to be predicted, and obtain the load forecasting result. Among them, Figure 3 - Figure 4 shows the performance comparison between the stacked model and the single model (composite predictor).

[0071] The technical effects of this application are as follows.

[0072] First, in the feature selection step, feature selection quantifies the degree of feature contribution, which largely determines the upper limit of the prediction accuracy of the corresponding task. By introducing a hybrid embedded algorithm, the representativeness and effectiveness of quantifying features before model establishment are improved.

[0073] Second, construct a sub-model pool through various heterogeneous artificial intelligence algorithms. This model framework establishes an advanced regression method to effectively exert its potential ability in the integration and improve the generalization and prediction performance of the model.

[0074] Third, the diversity regularization improvement of pruning the base learners in the sub-model pool realizes the optimal combination of individual base learners, effectively improving the prediction task performance and model performance.

[0075] Fourth, use a variety of heterogeneous supervised learning algorithms to establish a stacked ensemble load forecasting model, and perform data segmentation on the original data set based on the leave-one-out method to avoid overfitting under the stacked framework.

[0076] Fifth, optimize the prediction performance by expanding promising artificial intelligence algorithms, overcome the limitation that a single algorithm cannot reflect the true feature importance distribution, and improve the generalization ability of the load forecasting model.

[0077] Embodiment 2, this application also provides a computer system, which can be a server or a terminal, and its internal structure diagram can be as Figure 5 shown. The computer system includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer system is used to provide computing and control capabilities. The memory of the computer system includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer system is used to store processed data. The input / output interface of the computer system is used to exchange information between the processor and external devices. The communication interface of the computer system is used to communicate with external terminals through a network connection. When the computer program is executed by the processor, it implements a load forecasting method based on diversity regularization stacked learning.

[0078] Those skilled in the art can understand that Figure 5 the structure shown in Figure 5 is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer system to which the solution of this application is applied. The specific computer system may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0079] Embodiment 3, this application also provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the above-mentioned various methods are implemented.

[0080] Embodiment 4, this application also provides a computer program product including a computer program, and when the computer program is executed by a processor, the above-mentioned various methods are implemented.

[0081] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned method embodiments. Among them, any reference to a memory, database, or other medium used in the various embodiments provided in this application can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0082] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered to be within the scope described in this specification.

[0083] In this text, specific examples are used to elaborate on the principles and implementation manners of the present application. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation on the present application.

Claims

1. A load forecasting method based on diversity regularized stacking learning, characterized in that: The load forecasting method based on diversity regularized stacking learning includes: Acquire historical power operation data for training; the historical power operation data specifically includes: historical load, weather data and calendar rules; By using the embedded feature selection method and the correlation analysis method, the historical power operation data used for training is screened for relevance, and the screened historical power operation data is obtained; Performing diversity regularization pruning on the sub-model pool to obtain a target basic learner combination; the sub-model pool includes a plurality of basic learners for load forecasting; The target basic learner combination is stacked and trained according to the screened power operation history data to obtain a composite predictor; A composite predictor is used to perform load forecasting on the historical power operation data to be predicted to obtain the load forecasting result.

2. The load forecasting method based on diversity regularized stacking learning according to claim 1 is characterized in that: The embedded feature selection method and correlation analysis method are used to screen the historical power operation data used for training for relevance, and the screened historical power operation data are obtained, including: According to the historical power operation data used for training, a target predicted load contribution is obtained by using an embedded feature selection method; the target predicted load contribution includes: a historical load contribution, a weather data contribution, and a calendar rule contribution; Based on the target forecast load contribution, a correlation analysis method is used to analyze the correlation between the target forecast load and the historical load, weather data and calendar rules, respectively, to obtain the target forecast load correlation; the target forecast load correlation includes: the correlation of the historical load, the correlation of the weather data and the correlation of the calendar rules; The power operation history data used for training is screened according to the target predicted load relevance to obtain the screened power operation history data.

3. The load forecasting method based on diversity regularized stacking learning according to claim 1 is characterized in that: Perform diversity regularization pruning on the sub-model pool to obtain the target basic learner combination, including: Calculate the prediction generalization error of each base learner in the sub-model pool; Determine the individual accuracy of each base learner based on the prediction error calculation of each base learner; According to the individual accuracy of each base learner, the mutual information theory and hierarchical clustering are used to conduct regularized diversity analysis on the base learner combination to obtain the target base learner combination.

4. The load forecasting method based on diversity regularized stacking learning according to claim 3 is characterized in that: The calculation formula of the prediction generalization error of the base learner is as follows: Where e(l) is the generalization error of the base learner l, b 2 (l) is the bias of the base learner l, v(l) is the variance of the base learner l, μ 2 is the irreducible error, E(l) is the expected output of the base learner l, E(·) is the expected operator, Q(l) is the predicted value of the base learner l, and f is the true function.

5. The load forecasting method based on diversity regularized stacking learning according to claim 3 is characterized in that: According to the individual accuracy of each basic learner, the mutual information theory and hierarchical clustering are used to conduct regularized diversity analysis on the basic learner combination to obtain the target basic learner combination, which includes: The mutual information coefficient is obtained by using the mutual information theory analysis based on the individual accuracy of each basic learner; Using the mutual information coefficient, searching from the root of the tree diagram to determine the cluster that is most suitable for the basic learner pruning and diversity regularization; the tree diagram is a tree diagram obtained after the basic learner is hierarchically clustered according to the prediction error; The target base learner combination is selected and combined from multiple base learners according to the clustering that is most suitable for base learner pruning and diversity regularization.

6. The load forecasting method based on diversity regularized stacking learning according to claim 1, characterized in that: The target basic learner combination is stacked and trained according to the screened power operation history data to obtain a composite predictor, which includes: The screened power operation history data is divided into multiple data sets by using the leave-one-out method; each data set includes multiple first training sets and one test set; Using multiple first training sets in different data sets to perform stack training on the target base learner combination to obtain a trained target base learner; Input the test sets from different data sets into the trained target base learner to obtain multiple single model prediction results; Constructing a second training set based on multiple single model prediction results; The meta-learner is trained using the second training set to obtain a composite predictor.

7. The load forecasting method based on diversity regularized stacking learning according to claim 6 is characterized in that: The leave-one-out method is used to divide the screened power operation historical data into multiple data sets, including: The original training data is divided into multiple data blocks, and a different data block is selected each time as the test set, and the remaining data blocks are used as the first training set to obtain multiple data sets.

8. A computer system comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the load forecasting method based on diversity regularized stacked learning as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the load forecasting method based on diversity regularized stacked learning described in any one of claims 1 to 7 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the load forecasting method based on diversity regularized stacked learning described in any one of claims 1 to 7 is implemented.

Images