CN116011657B - Optimization method, device and system for power distribution network load prediction model based on miniature PMU - Google Patents

Abstract

The invention relates to the technical field of situation awareness, in particular to a power distribution network load prediction model optimization method, device and system based on a miniature PMU. The method comprises the steps of obtaining task data of a sample load prediction task and a feature set F of the sample load prediction task; training a plurality of load prediction models according to task data of sample load prediction tasks, and acquiring a corresponding optimal load prediction model for solving each sample load prediction task based on root mean square error to form an optimal load prediction model set phi; the feature set F and the optimal load prediction model set phi are combined to form metadata<F,Φ>The method comprises the steps of carrying out a first treatment on the surface of the Utilizing metadata<F,Φ>Training the plurality of meta learners respectively to obtain a plurality of trained meta learners; processing the load prediction task feature set by using a plurality of trained meta learners to obtain a plurality of model recommendation result data

Recommending multiple models to result data

Obtaining single model recommendation result data through a voter

Description

Distribution network load forecasting model optimization method, device and system based on micro PMU

Technical Field

The present invention relates to the field of situation awareness technology, and in particular to a method, device and system for optimizing a load forecasting model for a distribution network based on a micro PMU.

Background Art

Load forecasting of power systems is an important part of situational awareness and one of the basic supporting technologies for system dispatching and operation. Synchronous phasor measurement unit (PMU) is a phasor measurement unit that uses the second pulse of the GPS global positioning system as a synchronous clock. It can give a unified time scale to the global measurement data in the system, and can directly measure the phasor information in the distribution network. The synchronization and accuracy of the measurement data can be guaranteed, and it can be used in the fields of dynamic monitoring, system protection, system analysis and prediction of power systems. In recent years, due to the large-scale access of PMU to the power system, the operation and control of the power system have been greatly affected. How to make good use of the micro PMU in the power grid and study the key supporting technologies of the next generation of distribution network automation system based on the micro PMU technology that adapts to the needs of the distribution network has become a major scientific proposition in the field of energy and power systems.

With the large-scale access of distributed power sources, the current operation mode of the distribution network is becoming more and more flexible. For example, buildings equipped with photovoltaic and energy storage systems can achieve self-sufficiency in electricity as producers and sellers; some source networks in the system can be disconnected and operated independently as microgrids; the network topology of the system can be dynamically reconstructed as needed, etc. In order to meet these diverse operating needs, corresponding load forecasting technologies are needed to support them. Figure 1 describes different load forecasting tasks in the distribution network from five dimensions: data resolution, load level, influencing factors, forecast window, and historical data length. For example, the red pentagons are combined to define the day-ahead load forecasting task for a single user, with a data resolution of 1 hour, an available historical data length of 1 year, and temperature can be used as an influencing factor for modeling.

As can be seen from Figure 1, these heterogeneous prediction tasks have significant differences in data characteristics, prediction requirements, etc.

In 1995, D.H.Wolpert et al. proposed the No Free Lunch Theorem: Any prediction function that performs well on some training samples must perform poorly on other training samples; if there is no assumption about the prior distribution of data in the feature space, there will be as many cases of good performance as bad performance. According to the "No Free Lunch Theory", these heterogeneous prediction tasks are difficult to solve well with the same prediction model.

Existing Technology Long short-term memory (LSTM) is a special RNN that replaces the RNN cells in the hidden layer with LSTM cells to enable it to have long-term memory capabilities. After continuous evolution, the most widely used LSTM cell structure is shown in Figure 2. z is the input module, and its forward calculation method can be expressed as

i _t =σ(W _xi x _t +W _hi h _t-1 +W _ci c _t-1 +b _i ) (1)

f _t =σ(W _xf x _t +W _hf h _t-1 +W _cf c _t-1 +b _f ) (2)

c _t ＝f _t c _t-1 +i _t tanh(W _xc x _t +W _hc h _t-1 +b _c ) (3)

o _t =σ(W _xo x _t +W _ho h _t-1 +W _co c _t +b _o ) (4)

h _t = o _t tanh(c _t ) (5)

Where i, f, c, and o are the input gate, forget gate, cell state, and output gate, respectively; W and b are the corresponding weight coefficient matrix and bias term, respectively; σ and tanh are the sigmoid and hyperbolic tangent activation functions, respectively.

The LSTM model training process can be roughly divided into four steps: calculate the output value of the LSTM cell according to the forward calculation method of equations (1) to (5); reversely calculate the error term of each LSTM cell, including two back propagation directions according to time and network level; calculate the gradient of each weight according to the corresponding error term; and apply the gradient-based optimization algorithm to update the weight.

LSTM networks are very suitable for classification, processing and prediction based on time series data, but they are not necessarily suitable for all load forecasting tasks in distribution networks. This is because the LSTM model requires a large number of training samples and a long calculation time, which makes it difficult to apply to forecasting scenarios with small samples and high real-time requirements. In addition, the historical load series of some forecasting tasks are relatively smooth and can be solved by a simpler forecasting model without the introduction of the LSTM model. The above examples show that LSTM is not a universal forecasting method that can solve heterogeneous forecasting tasks in distribution networks. The model optimization architecture proposed in this article can select the best forecasting model for different forecasting tasks, which better solves the model adaptation problem.

Summary of the invention

In order to solve the technical problems existing in the above-mentioned prior art, the present invention provides a method, device and system for optimizing a distribution network load prediction model based on a micro PMU.

To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

In a first aspect, in one embodiment provided by the present invention, a method for optimizing a distribution network load forecasting model based on a micro PMU is provided, the method comprising the following steps:

Obtaining task data of a sample load forecasting task and a feature set F of the sample load forecasting task;

Training multiple load forecasting models according to the task data of the sample load forecasting task, and obtaining the corresponding optimal load forecasting model for solving each sample load forecasting task based on the root mean square error to form an optimal load forecasting model set Φ;

The feature set F and the optimal load forecasting model set Φ together constitute metadata <F,Φ>; using the metadata <F,Φ> to train multiple meta-learners respectively to obtain multiple trained meta-learners;

将所述多个模型推荐结果数据

通过投票器，获得单一模型推荐结果数据

Use multiple trained meta-learners to process the load forecasting task feature set and obtain multiple model recommendation result data

The plurality of model recommendation result data

Obtain single model recommendation result data through the voting machine

As a further solution of the present invention, the sample load prediction task data includes J data sample pairs <X _j ,y _j >, where X _j is the input data of the load prediction model with a dimension of N _j ×M _j ; y _j is the actual load value with a dimension of N _j ×1, where j∈[1,..,J].

和

y_j包括

和

As a further embodiment of the present invention, the _Xj comprises

and

y _j includes

and

通过如下公式进行计算：Among them, the

Calculate using the following formula:

表示第n个样本的第m个属性。针对第j个负荷预测任务，将负荷预测模型i_B生成的预测结果记为

in

represents the mth attribute of the nth sample. For the jth load forecasting task, the forecast result generated by the load forecasting model i _B is recorded as

As a further solution of the present invention, the dimension of the feature set F is J×D.

As a further solution of the present invention, the load forecasting model is calculated by the following formula:

为衡量预测结果和真实结果之间距离的损失函数，θ^*则是负荷预测模型i_B的最优参数集合。in

is the loss function that measures the distance between the predicted result and the actual result, and θ ^* is the optimal parameter set of the load forecasting model i _B.

As a further solution of the present invention, the metadata <F, Φ> is divided into a metadata training set <F ^train , Φ ^train > and a metadata test set <F ^test , Φ ^test >.

As a further solution of the present invention, the step of training a plurality of meta-learners using the metadata <F, Φ> to obtain trained meta-learners further includes:

During the training process, multiple meta-learners generate corresponding multiple training model recommendation result data

通过投票器，获得单一训练模型推荐结果数据

The plurality of training model recommendation result data

Obtain the recommendation result data of a single training model through the voting machine

In a second aspect, in another embodiment provided by the present invention, a distribution network load forecasting model optimization device based on a micro PMU is provided, the device comprising: a data acquisition module, a first training module, a second training module and an application module;

The data acquisition module is used to acquire task data of a sample load prediction task and a feature set F of the sample load prediction task, wherein the sample load prediction task includes J sample load prediction tasks;

The first training module is used to train multiple load forecasting models according to the task data of the sample load forecasting task, and obtain the corresponding optimal load forecasting model for solving each sample load forecasting task based on the root mean square error to form an optimal load forecasting model set Φ;

The second training module is used to form metadata <F, Φ> by combining the feature set F and the optimal load forecasting model set Φ, and train multiple meta-learners respectively by using the metadata <F, Φ> to obtain multiple trained meta-learners;

将所述多个模型推荐结果数据

通过投票器，获得单一模型推荐结果数据

The application module is used to use multiple trained meta-learners to process the load forecasting task feature set respectively to obtain multiple model recommendation result data

The plurality of model recommendation result data

Obtain single model recommendation result data through the voting machine

In a third aspect, in another embodiment provided by the present invention, a distribution network load forecasting model optimization system based on a micro PMU is provided, the system comprising: a basic learning layer, a meta-learning layer, and an application layer;

The basic learning layer is used to obtain task data of a sample load forecasting task, train multiple load forecasting models according to the task data of the sample load forecasting task, and obtain the corresponding optimal load forecasting model for solving each sample load forecasting task based on a root mean square error to form an optimal load forecasting model set Φ;

The meta-learning layer is used to obtain a feature set F of a sample load forecasting task, and to form metadata <F, Φ> with the feature set F and the optimal load forecasting model set Φ; and to train a plurality of meta-learners respectively using the metadata <F, Φ> to obtain a plurality of trained meta-learners;

将所述多个模型推荐结果数据

通过投票器，获得单一模型推荐结果数据

The application layer is used to use multiple trained meta-learners to process the load forecasting task feature set respectively to obtain multiple model recommendation result data

The plurality of model recommendation result data

Obtain single model recommendation result data through the voting machine

The technical solution provided by the present invention has the following beneficial effects:

将所述多个模型推荐结果数据

通过投票器，获得单一模型推荐结果数据

本发明采用元学习技术，对配电网中各个微型同步相量测量单元(Phasor Measurement Unit,PMU)推荐最优负荷预测模型，满足配电网中异质的负荷预测任务的要求，提升整体的预测精度。The present invention provides a method, device and system for optimizing a load forecasting model for a distribution network based on a micro PMU. The present invention obtains task data of a sample load forecasting task and a feature set F of the sample load forecasting task; trains multiple load forecasting models according to the task data of the sample load forecasting task, and obtains the corresponding optimal load forecasting model for solving each sample load forecasting task based on the root mean square error to form an optimal load forecasting model set Φ; the feature set F and the optimal load forecasting model set Φ together constitute metadata <F, Φ>; uses the metadata <F, Φ> to train multiple meta-learners respectively to obtain multiple trained meta-learners; uses the multiple trained meta-learners to process the load forecasting task feature set respectively to obtain multiple model recommendation result data

The plurality of model recommendation result data

Obtain single model recommendation result data through the voting machine

The present invention adopts meta-learning technology to recommend the optimal load forecasting model to each micro-synchronous phasor measurement unit (PMU) in the distribution network, so as to meet the requirements of heterogeneous load forecasting tasks in the distribution network and improve the overall forecasting accuracy.

These and other aspects of the present invention will become more concise and understandable in the following description of the embodiments. It should be understood that the above general description and the following detailed description are only exemplary and explanatory and cannot limit the present invention.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other embodiments can be obtained based on these drawings without paying creative work.

Figure 1 shows the heterogeneous load forecasting task in the distribution network.

Figure 2 is a diagram of the LSTM hidden layer cell structure.

FIG3 is a flow chart of a method for optimizing a distribution network load forecasting model based on a micro PMU according to an embodiment of the present invention.

FIG4 is a structural block diagram of a distribution network load forecasting model optimization device based on a micro PMU according to an embodiment of the present invention.

FIG5 is a structural block diagram of a distribution network load forecasting model optimization system based on a micro PMU according to an embodiment of the present invention.

FIG6 is a voting process of a meta-learner in a distribution network load forecasting model optimization system based on a micro-PMU according to an embodiment of the present invention.

Figure 7 shows an example of the model selection process.

Figure 8 shows the model labeling results.

FIG9 shows the SER ratios of different prediction models.

FIG10 is a diagram showing the dimensionality reduction processing of the feature set based on random proximity coding of T distribution.

Figure 11 shows the test results of the meta-learner in cluster 4.

Figure 12 shows the prediction result of the recommended model a.

Figure 13 shows the prediction results of the recommendation model b.

In the figure: data acquisition module-100, first training module-200, second training module-300, application module-400.

DETAILED DESCRIPTION

The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations/steps, nor must they be executed in the order described. For example, some operations/steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

It should be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

Specifically, the embodiments of the present invention are further described below in conjunction with the accompanying drawings.

Please refer to FIG. 3 , which is a flow chart of a method for optimizing a distribution network load forecasting model based on a micro PMU provided in an embodiment of the present invention. As shown in FIG. 3 , the method for optimizing a distribution network load forecasting model based on a micro PMU includes steps S10 to S40.

S10. Obtain task data of a sample load prediction task and a feature set F of the sample load prediction task, wherein the sample load prediction task includes J sample load prediction tasks.

In an embodiment of the present invention, the sample load forecasting task data includes J data sample pairs <X _j ,y _j >, where X _j is the input data of the load forecasting model, and its dimension is N _j ×M _j ; y _j is the actual load value, and its dimension is N _j ×1, where j∈[1,..,J]. The value range of J is a positive integer.

和

y_j包括

和

In the embodiment of the present invention, in order to train the load forecasting model, _Xj includes

and

y _j includes

and

通过如下公式进行计算：Among them, the

Calculate using the following formula:

表示第n个样本的第m个属性。针对第j个负荷预测任务，将负荷预测模型i_B生成的预测结果记为

in

represents the mth attribute of the nth sample. For the jth load forecasting task, the forecast result generated by the load forecasting model i _B is recorded as

S20. Train multiple load forecasting models according to the task data of the sample load forecasting task, and obtain the corresponding optimal load forecasting model for solving each sample load forecasting task based on the root mean square error to form an optimal load forecasting model set Φ.

The number of load forecasting models is I _B. The value of I _B can be flexibly configured according to the actual needs of users. It should be noted that the selected I _B models should have different advantages to complement each other, so as to solve diverse forecasting tasks.

The load forecasting model is calculated using the following formula:

为衡量预测结果和真实结果之间距离的损失函数，θ为预测模型的待定参数，θ^*则是负荷预测模型i_B的最优参数集合。当通过训练得到θ^*后，则可通过计算在测试集上的RMSE误差来衡量预测精度，即in

is a loss function that measures the distance between the predicted result and the actual result. θ is the undetermined parameter of the prediction model, and θ ^* is the optimal parameter set of the load prediction model i _B. After θ ^* is obtained through training, the prediction accuracy can be measured by calculating the RMSE error on the test set, that is,

Among all I _B candidate load forecasting models, the model with the highest prediction accuracy for load forecasting task j will be labeled as Φ(j).

S30, the feature set F and the optimal load prediction model set Φ together form metadata <F, Φ>; use the metadata <F, Φ> to train multiple meta-learners respectively to obtain multiple trained meta-learners.

The dimension of the feature set F is J×D, where D is the number of features in the feature set.

In order to train the meta-learner, the metadata <F, Φ> is divided into a metadata training set <F ^train , Φ ^train > and a metadata test set <F ^test , Φ ^test >.

The meta-learner performs the following computation:

是元学习器的损失函数，用以衡量所推荐模型

与实际最优模型Φ^train之间的距离。将w经过训练之后所得到的最优参数记为w^*。当元学习器通过训练获得其最优参数w^*之后，可在测试集上对其精度进行测试，推荐精度η_iM通过如下公式进行计算：Where _gw represents the meta-learner, w is the parameter to be trained for meta-learning,

is the loss function of the meta-learner, which is used to measure the recommended model

The distance between the actual optimal model Φ ^train . The optimal parameter obtained after training is recorded as w ^* . After the meta-learner obtains its optimal parameter w ^* through training, its accuracy can be tested on the test set, and the recommended accuracy η _iM is calculated by the following formula:

Among them, K _M is the number of load forecasting tasks in meta-learning training.

Specifically, the meta-learner is essentially a multi-classification machine learning model whose parameters are w (for example, if the meta-learner is a neural network, then w here refers to the weights of each neuron in the network). After the model is trained using data, w will slowly converge to the optimal value, that is, w ^* .

通过本发明设计得投票器，可综合以上推荐结果并获得最终的推荐结果

相应的推荐精度为η。Since multiple meta-learners are used in the present invention, multiple model recommendation results will be generated.

The voting device designed by the present invention can synthesize the above recommendation results and obtain the final recommendation result.

The corresponding recommended accuracy is η.

In an embodiment of the present invention, the step of training a plurality of meta-learners using the metadata <F, Φ> to obtain trained meta-learners further includes:

During the training process, the multiple meta-learners generate corresponding multiple training model recommendation result data

通过投票器，获得单一训练模型推荐结果数据

The plurality of training model recommendation result data

Obtain the recommendation result data of a single training model through the voting machine

The number of the meta-learners is _IM , and its value can be flexibly configured according to the actual needs of the user, and the minimum value is 1.

将所述多个模型推荐结果数据

通过投票器，获得单一模型推荐结果数据

S40, using multiple trained meta-learners to process the load forecasting task feature set respectively, and obtain multiple model recommendation result data

The plurality of model recommendation result data

Obtain single model recommendation result data through the voting machine

The present invention adopts meta-learning technology to recommend the optimal load forecasting model to each micro-synchronous phasor measurement unit (PMU) in the distribution network, so as to meet the requirements of heterogeneous load forecasting tasks in the distribution network and improve the overall prediction accuracy.

It should be understood that, although described in a certain order, these steps are not necessarily performed in sequence in the above order. Unless there is clear explanation in this article, the execution of these steps does not have strict order restriction, and these steps can be performed in other orders. Moreover, a part of the steps of the present embodiment may include a plurality of steps or a plurality of stages, and these steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a part of the steps or stages in other steps or other steps.

In one embodiment, as shown in FIG. 4 , a distribution network load forecasting model optimization device based on a micro PMU is also provided in an embodiment of the present invention. The device includes a data acquisition module 100 , a first training module 200 , a second training module 300 and an application module 400 .

The data acquisition module 100 is used to acquire task data of a sample load prediction task and a feature set F of the sample load prediction task, wherein the sample load prediction task includes J sample load prediction tasks.

The first training module 200 is used to train multiple load forecasting models according to the task data of the sample load forecasting task, and obtain the corresponding optimal load forecasting model for solving each sample load forecasting task based on the root mean square error to form an optimal load forecasting model set Φ.

The second training module 300 is used to form metadata <F, Φ> by combining the feature set F and the optimal load prediction model set Φ, and use the metadata <F, Φ> to train multiple meta-learners respectively to obtain multiple trained meta-learners.

将所述多个模型推荐结果数据

通过投票器，获得单一模型推荐结果数据

The application module 400 is used to process the load forecasting task feature set using multiple trained meta-learners to obtain multiple model recommendation result data

The plurality of model recommendation result data

Obtain single model recommendation result data through the voting machine

The present invention adopts meta-learning technology to recommend the optimal load forecasting model to each micro-synchronous phasor measurement unit (PMU) in the distribution network, so as to meet the requirements of heterogeneous load forecasting tasks in the distribution network and improve the overall prediction accuracy.

In one embodiment, as shown in FIG5 , a distribution network load forecasting model optimization system based on a micro PMU is also provided in an embodiment of the present invention. The system includes a basic learning layer, a meta-learning layer, and an application layer.

The basic learning layer is used to obtain task data of sample load forecasting tasks, train multiple load forecasting models according to the task data of the sample load forecasting tasks, and obtain the corresponding optimal load forecasting model for solving each sample load forecasting task based on the root mean square error to form an optimal load forecasting model set Φ.

The meta-learning layer is used to obtain a feature set F for a sample load forecasting task, and to form metadata <F,Φ> with the feature set F and the optimal load forecasting model set Φ; and to train multiple meta-learners respectively using the metadata <F,Φ> to obtain multiple trained meta-learners.

将所述多个模型推荐结果数据

通过投票器，获得单一模型推荐结果数据

The application layer is used to use multiple trained meta-learners to process the load forecasting task feature set respectively to obtain multiple model recommendation result data

The plurality of model recommendation result data

Obtain single model recommendation result data through the voting machine

The present invention adopts meta-learning technology to recommend the optimal load forecasting model to each micro-synchronous phasor measurement unit (PMU) in the distribution network, so as to meet the requirements of heterogeneous load forecasting tasks in the distribution network and improve the overall prediction accuracy.

For example, please provide an embodiment of a specific application process of a distribution network load forecasting model optimization system based on a micro PMU, given a load forecasting model, a load forecasting task feature set and a load forecasting task, the steps are as follows:

Step S100: Create a sample load forecasting task.

As shown in Figure 1, a large number of heterogeneous sample load forecasting tasks will be created from five dimensions.

Step S200, select four commonly used load forecasting models as alternatives. The load forecasting models include: Autoregressive Integrated Moving Average (SARIMA) model with seasonal characteristics, LSTM model, Support Vector Regression (SVR), and Similar Day (SD) model. The four load forecasting models represent four different ideas of time series prediction, deep learning prediction, correlation factor prediction, and cluster prediction. Six different parameter structures are set for the SARIM model, and two different parameter structures are set for the LSTM model to verify that the meta-learner can not only select a large model category, but also select the optimal model structure. Finally, 10 candidate models are obtained as shown in Table 1.

Table 1 Alternative load forecasting models

Table 1 Candidate load forecasting models

Step S300: Load forecasting model marking.

In order to find the most suitable prediction model Φ(j) for load forecasting task j, formulas (7)-(10) will be repeated Lj times, each time using different training and test data segmentation, so as to obtain the distribution Ω _j of the optimal load forecasting model. This process continues until the distribution Ω _j tends to be stable. The Pearson correlation coefficient P _cc can be used as a termination condition, that is, when P _cc between Ω _j (L _j ) and Ω _j (L _j -10) is greater than 0.95, the distribution is considered to have reached a stable state. For details of the algorithm, see Algorithm 1.

Step S400: Design a feature set containing 16 features to quantitatively describe the load forecasting task, as shown in Table 2.

Table 2 Load forecasting task feature set

Table2 Feature set of forecasting tasks

为历史负荷序列的标准差和均值，y(n)代表数据序列第n个值。Kurtosis and skewness can be calculated by equations (16) and (17), where y is the historical load data sequence, N is the length of the data sequence, σ,

are the standard deviation and mean of the historical load series, and y(n) represents the nth value of the data series.

Volatility (Fickleness) measures the number of times a historical series crosses its mean line

The correlation characteristics of the maximum autocorrelation coefficient (H-ACF) and the maximum partial autocorrelation coefficient (H-PACF) load series are very important for determining the structure of the SARIMA model. The period is directly related to the resolution of the data. For example, for a load series with an hourly resolution, the period is mostly 24 or 168. For a series with a daily resolution, the period is mostly 30.

Step S500: Meta-learner selection and voting.

Bayesian,NB)，以及线性判别分析(Linear Discrimination,LD)。为了结合各个分类器的推荐结果，并获得最终的推荐模型，本发明基于集成学习的思想设计了投票器，如图6所示。The meta-learner in the present invention realizes the mapping from task feature F to optimal load forecasting model Φ(j), which is essentially a multi-classification problem. In order to improve the classification accuracy of the meta-learner, the meta-learner in this system includes 4 different classifiers. The classifiers include Random Forest (RF), K-Nearest Neighbor (KNN), Naive Bayes (

In order to combine the recommendation results of each classifier and obtain the final recommendation model, the present invention designs a voting machine based on the idea of ensemble learning, as shown in FIG6 .

与分类精度

之间的映射关系The mechanism by which each meta-learner achieves classification is based on an internal scoring process. For example, NB calculates the posterior probability of each class as its score, while RF calculates the voting results of the decision trees contained internally as its score. The meta-learner iM selects its highest-scoring model as its output. A higher score means that the meta-learner has a higher confidence in its recommendation results. Based on this feature, we establish a scoring

And classification accuracy

The mapping relationship between

When recommending models for a new load forecasting task, the score of each meta-learner is first converted into its corresponding recommendation accuracy according to the above formula, and on this basis, the model with the highest recommendation accuracy is selected as the final model recommendation.

result

Step S600: Create a load forecasting task.

Creating a large number of real and heterogeneous load forecasting tasks is the basis for testing the effectiveness of meta-learning. As shown in Table 3, we combine the five dimensions shown in Figure 1 to create different load forecasting tasks.

Table 3 Heterogeneous forecasting task creation

Residential and commercial load data with 15-minute and 30-minute accuracy are collected from North Carolina, and data with 1-minute accuracy are collected from Pecan Street. Weather data with hourly accuracy are collected from the NOAA website. By combining the five dimensional features in Table 3, this paper obtains a total of 846 different load forecasting tasks. These tasks have the following characteristics: 1) They mainly include residential and commercial loads, and industrial and agricultural loads are not considered; 2) From a single user to a microgrid, the most commonly used short-term load forecasting is mainly considered to support its operation; 3) At the feeder level, medium- and long-term load forecasting tasks are additionally considered; 4) Up to 12 weather features can be obtained; 5) Three different historical data lengths are considered to reflect the differences in the availability of historical data in actual engineering applications.

Step S700: The basic learning layer obtains the model optimization result.

For these 846 load forecasting tasks, the optimal forecasting model is found and marked. Figure 7 shows the model selection process for a load forecasting task as an example. As the number of iterations increases, the distribution of the optimal model gradually stabilizes. At 60 iterations, the correlation coefficient between its frequency distribution and that of the optimal model at 50 iterations reaches 0.98>0.95, so the iteration is terminated and model 9 with the maximum frequency is selected as the optimal model for the load forecasting task.

The model labeling results for all 846 load forecasting tasks are shown in Figure 8. It should be noted that if the historical data of a load forecasting task is insufficient to train the model, the model is considered to have failed and a significant prediction error is assigned to the model to prevent it from being selected.

From the results, we can see that model 7, LSTM (125), is the most frequently selected model. In addition, model 10, the similar day model, has the shortest training time, but the mean and variance of its prediction error are larger than those of other models. Since the SARIMA model has high requirements for historical data, the number of training failures is relatively large, especially for high-order SARIMA models. However, when the SARIMA model can be trained, it has a relatively good prediction performance.

In order to further quantify the differences between different prediction models, the system error ratio (SER) is defined

Where Eselect is the RMSE of the recommended model, and Ebest is the RMSE of the optimal model. This indicator measures the distance between the recommended model and the actual optimal model. Figure 9 shows the performance differences of different models in the load forecasting task. It can be seen that the models ranked 2-4 have similar performance to the optimal model in most cases, that is, the SER is close to 1. However, the performance of the models ranked lower is significantly worse than the optimal model. This shows the necessity of carrying out model selection work.

Step S800: Load forecasting task similarity evaluation.

In the meta-learning layer, the input of the meta-learner is the feature set F of the load forecasting task. We apply T-distributed Stochastic Neighbor Embedding (t-SNE) to reduce the dimension of the feature set and plot it on a two-dimensional plane as shown in Figure 10. As can be seen from the figure, the load forecasting tasks are roughly clustered into 5 clusters, as shown in Table 4.

Table 4 Load forecasting task cluster characteristics

Table 4 Features of each LF cluster

Clusters 1, 2, and 5 represent load forecasting tasks at the feeder level. Cluster 1 represents medium- and long-term load forecasting tasks in days, and the best model is SD. This is because there is less historical data for this type of task, and complex models are difficult to train well, while the simple SD model is sufficient to provide better forecasting results. Clusters 2 and 5 represent short-term load forecasting tasks with weather characteristics, and the best model is LSTM. Cluster 3 represents short-term load forecasting tasks at the user and transformer levels, and SVR is its optimal solution. Cluster 4 represents short-term load forecasting tasks at the microgrid level, and the best model is SARIMA.

Step S900: training and verification results of the meta-learner.

In order to train and test the meta-learner, we divide the 846 load forecasting tasks into a training set (70%), a validation set (20%), and a test set (10%). Figure 11 shows the test results of the trained meta-learner on cluster 4. It can be seen that different meta-learners have their own successful and failed tasks. Effectively combining them can improve the overall model optimization accuracy, as shown in Table 5.

Table 5 Meta-learner test accuracy

Table 5 Accuracy of meta learners

While recommending the optimal model, the recommendation system established in this paper can give the ranking of all candidate models for each load forecasting task, thus providing a reference for operators. As shown in Table 6, it can be seen that the top 3 models all have a high recommendation success rate and a low SER error ratio, and can all be used as effective models. The probability of the recommendation system recommending an effective model is 76%.

Table 6 Comparison of accuracy of different sorting models

Table6 Accuracy comparison of LF models under different rankings

Sorting 1 2 3 4 5 6 7 8 9 10 Accuracy 46% 17% 13% 6% 4% 3% 3% 3% 2% 3% SER 1.14 1.27 1.34 1.46 4.18 2.89 4.48 3.61 2.61 3.09 Number of failures 0 0 2 10 10 12 12 17 14 11

Step S1000: Online application test.

After the recommendation system is trained, it can be applied online to recommend prediction models for new load forecasting tasks. This section considers model recommendation tests for two new load forecasting tasks, and the task descriptions are shown in Table 7. Task 1 is short-term load forecasting at the transformer level, with historical data of 30 days and 15 minutes resolution; Task 2 is short-term load forecasting at the feeder level, with historical data of 6 months and 1 hour resolution. According to the recommendation system, the recommended model for Task 1 is SD, which is the true optimal model; the recommended model for Task 2 is SARIMA (5,1,5), which is the true second-best model. The recommended model is used for prediction, and the results are shown in Figures 12 and 13.

Table 7 Two new test tasks

Table 7 Two testing LF tasks

It should be further understood that the term “and/or” used in the present specification and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

It should be understood that, as used herein, the singular form "a" or "an" is intended to include the plural form as well, unless the context clearly supports an exception. It should also be understood that, as used herein, "and/or" refers to any and all possible combinations of one or more of the items listed in association. The serial numbers of the embodiments disclosed in the above embodiments of the present invention are for description only and do not represent the advantages and disadvantages of the embodiments.

A person skilled in the art should understand that the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope of the disclosure of the embodiments of the present invention (including the claims) is limited to these examples; under the idea of the embodiments of the present invention, the technical features in the above embodiments or different embodiments can also be combined, and there are many other changes in different aspects of the above embodiments of the present invention, which are not provided in detail for the sake of simplicity. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present invention should be included in the protection scope of the embodiments of the present invention.

Claims (3)

1. A method for optimizing a distribution network load forecasting model based on a micro PMU, characterized in that the method comprises: Obtaining task data of a sample load forecasting task and a feature set F of the sample load forecasting task; Training multiple load forecasting models according to the task data of the sample load forecasting task, and obtaining the corresponding optimal load forecasting model for solving each sample load forecasting task based on the root mean square error to form an optimal load forecasting model set Φ; The feature set F and the optimal load forecasting model set Φ together constitute metadata <F, Φ>; use the metadata <F, Φ> to train multiple meta-learners respectively to obtain multiple trained meta-learners; Use multiple trained meta-learners to process the load forecasting task feature set and obtain multiple model recommendation result data

The plurality of model recommendation result data

Obtain single model recommendation result data through the voting machine

The sample load forecasting task data includes J data sample pairs <X _j ,y _j >, where X _j is the input data of the load forecasting model, and its dimension is N _j ×M _j ; y _j is the real load value, and its dimension is N _j ×1, where j∈[1,..,J]; The _Xj includes

and

y _j includes

and

Among them, the

Calculated using the following formula:

in

represents the mth attribute of the nth sample; for the jth load forecasting task, the forecast result generated by the load forecasting model i _B is recorded as

The dimension of the feature set F is J×D; The load forecasting model is calculated using the following formula:

in

is the loss function that measures the distance between the predicted result and the true result, and θ ^* is the optimal parameter set of the load forecasting model i _B ; The metadata <F, Φ> is divided into a metadata training set <F ^train , Φ ^train > and a metadata test set <F ^test , Φ ^test >; The step of training a plurality of meta-learners using the metadata <F, Φ> to obtain trained meta-learners further includes: During the training process, the multiple meta-learners generate corresponding multiple training model recommendation result data

The plurality of training model recommendation result data

Obtain the recommendation result data of a single training model through the voting machine

2. A device for optimizing a load forecasting model for a distribution network based on a micro PMU using the method of claim 1, characterized in that the device comprises: a data acquisition module, a first training module, a second training module and an application module; The data acquisition module is used to acquire task data of a sample load prediction task and a feature set F of the sample load prediction task, wherein the sample load prediction task includes J sample load prediction tasks; The first training module is used to train multiple load forecasting models according to the task data of the sample load forecasting task, and obtain the corresponding optimal load forecasting model for solving each sample load forecasting task based on the root mean square error to form an optimal load forecasting model set Φ; The second training module is used to form metadata <F, Φ> by combining the feature set F and the optimal load forecasting model set Φ, and train multiple meta-learners respectively by using the metadata <F, Φ> to obtain multiple trained meta-learners; The application module is used to use multiple trained meta-learners to process the load forecasting task feature set respectively to obtain multiple model recommendation result data

The plurality of model recommendation result data

Obtain single model recommendation result data through the voting machine

3. A distribution network load forecasting model optimization system based on micro PMU using the method described in claim 1, characterized in that the system comprises: a basic learning layer, a meta-learning layer, and an application layer; The basic learning layer is used to obtain task data of a sample load forecasting task, train multiple load forecasting models according to the task data of the sample load forecasting task, and obtain the corresponding optimal load forecasting model for solving each sample load forecasting task based on a root mean square error to form an optimal load forecasting model set Φ; The meta-learning layer is used to obtain a feature set F of a sample load forecasting task, and to form metadata <F,Φ> with the feature set F and the optimal load forecasting model set Φ; and to train a plurality of meta-learners respectively using the metadata <F,Φ> to obtain a plurality of trained meta-learners; The application layer is used to use multiple trained meta-learners to process the load forecasting task feature set respectively to obtain multiple model recommendation result data

The plurality of model recommendation result data

Obtain single model recommendation result data through the voting machine

Images