CN115660038A - Multi-stage integrated short-term load prediction based on error factors and improved MOEA/D-SAS - Google Patents

Abstract

The invention discloses a multi-stage integration method based on an error factor and a multi-objective evolution algorithm (MOEA \ D-SAS) with sorting and selecting functions and based on decomposition, which is used for solving the problem of short-term load prediction. The method comprises three stages: in the first stage, a gated cyclic unit (GRU) is adopted to predict a component of complete set empirical mode decomposition (CEEMDAN) decomposition of adaptive noise, and the component is combined with an original data set to obtain a new data set so as to fully mine data characteristics; in the second stage, the improved MOEA \ D-SAS is adopted, and the GRU network parameters are optimized by taking the accuracy and the diversity as objective functions, so that a plurality of load prediction models and error prediction models which take the accuracy and the diversity into consideration are obtained; and in the third stage, nonlinear integration is carried out on the load predicted value and the error predicted value obtained in the second stage, and error factors are considered to further improve the prediction precision. The method can better fit complex power load data, improves the accuracy of the prediction model, ensures the generalization capability of the prediction model, realizes accurate prediction on short-term power load, and provides data support for power dispatching.

Description

Multi-stage integrated short-term load prediction based on error factors and improved MOEA/D-SAS

Technical Field

The invention relates to the technical field of short-term power load prediction, in particular to an error factor-based multi-target GRU integrated network for short-term load prediction.

Background

With the increase of the population of the world and the improvement of the living standard of human beings, the energy consumption is higher and higher. Recently, energy consumption is mainly based on traditional fossil energy, resulting in a large amount of CO ₂ The greenhouse effect is increasingly obvious, and great challenges are generated to the living environment of human beings. China gradually turns to an economic development model based on new energy. This makes the electric load fluctuation increase day by day, and the load characteristic is more and more complicated, has increased the risk of grid safe operation. Accurate short-term load prediction is beneficial to maintaining normal operation of the power system, is an indispensable part of energy system management of the smart grid, and has important reference functions on real-time power dispatching, power system reliability improvement, operation cost reduction and market income influence.

With the continuous breakthrough of computer technology, the deep learning technology has stronger function approximation capability, and is widely applied to large-scale power load data information mining. LSTM, a typical representative of deep learning, has been a research focus in the field of load prediction due to its unique network structure and advantage of processing timing information. The GRU is a variant of the LSTM, not only inherits the advantages of the LSTM, but also has a simple structure and fewer parameters, and becomes an effective method for load prediction. However, in the face of increasingly complex load data, it is difficult for a single prediction model to achieve ideal prediction accuracy. For this reason, the integrated network is applied in deep learning. The integrated network can integrate the advantages of different submodels and is superior to a single model in the aspects of prediction accuracy and generalization. However, how to select an appropriate sub-model, improve the performance of the sub-model, and integrate the advantages of the sub-model is a challenging task.

To improve prediction performance, one combines optimization algorithms with prediction models. However, the sub-model optimization algorithm based on the single target focuses on improving accuracy, so that the model is easy to generate an overfitting phenomenon. Meanwhile, due to the problem of the multi-objective optimization algorithm, an abnormal situation is generated when the actual problem is processed. Taking the MOEA \ D algorithm as an example, when the parameters of the algorithm are set due to factors such as large load data volume, long neural network training time and the like, the population cannot be too large, and the iteration times are not too many, so that: (1) The iteration time is limited, and the MOEA/D neighborhood updating strategy is easy to generate a local optimal phenomenon in response to the occurrence of super individual conditions; (2) The population scale is too small, and a single population generation strategy is difficult to effectively explore a decision space in a limited time. The obtained submodel has the conditions of poor performance and similar models, and the integrated output is seriously influenced. In addition, the output result of the traditional integration strategy is biased to the optimal submodel performance or the average performance of the submodels, the advantages of different submodels are neglected, and the requirement of an integration network is difficult to meet.

Disclosure of Invention

Aiming at the technical problem of the short-term load prediction method based on the integrated model, the invention discloses a GRU multi-stage integrated model for short-term load prediction based on error factors and improved MOEA/D-SAS. The model is divided into three stages: in the first stage, components decomposed by CEEMDAN are predicted by adopting GRU, and are combined with an original data set to obtain a new data set so as to fully mine data characteristics; in the second stage, the improved MOEA \ D is adopted to optimize GRU network parameters by taking accuracy and diversity as objective functions so as to obtain a plurality of MOGRU load prediction models and error prediction models which take the accuracy and the diversity into consideration; and in the third stage, nonlinear integration is carried out on the load predicted value and the error predicted value obtained in the second stage, and error factors are considered to further improve the prediction precision.

The scheme for solving the technical problems is as follows:

generating and optimizing a sub-model through an improved MOEA/D-SAS optimization algorithm;

removing unsuitable individuals in the population through an optimal selection strategy;

establishing an error prediction model for each load prediction model;

integrating and outputting a load prediction model and an error prediction model by a MOGRU-based nonlinear integration method;

the invention has the technical effects that:

according to the method, the GRU is adopted to predict the CEEMDAN decomposed component, and the component is combined with the original data set to obtain a new data set, so that the prediction complexity of load data is reduced, and data characteristics are fully mined; the learning rate of the GRU network and the number of hidden nodes of each layer are used as a decision space, the accuracy and diversity of the GRU network are used as a target space, the MOEA/D-SAS optimization algorithm is used for optimizing the submodel, and the accuracy of the submodel is improved while the diversity of the submodel is ensured; selecting the optimal selection strategy to eliminate the individual with problems in the optimization result, and providing the most appropriate sub-model for integrated output; a MOEA \ D-SAS-based sub-model generation strategy is used for establishing a load prediction model and an error prediction model, and the prediction accuracy is further improved by correcting errors of a load prediction value; and (3) performing integrated output on the submodels by using a MOGRU-based nonlinear integration method, and extracting the advantages of each submodel and avoiding the disadvantages of each submodel through the strong nonlinear learning capability of the GRU network.

Drawings

FIG. 1 is a schematic diagram of a MOEA/D-SAS-based sub-model generation strategy in the invention.

Fig. 2 is a schematic diagram of a MOGRU-based nonlinear integration strategy according to the present invention.

Fig. 3 is an overall flowchart in the present invention.

Detailed Description

The present invention will be described in further detail with reference to the accompanying drawings.

The method comprises the following implementation steps:

based on an improved MOEA \ D-SAS sub-model generation strategy:

step 1: constructing a decision space of the MOEA/D-SAS optimization algorithm by using the learning rate of the GRU network and the number of hidden nodes of each layer, and using the accuracy and diversity of the GRU network as a target space of the MOEA/D-SAS optimization algorithm:

x＝{x ₁ ,x ₂ ,}#(1)

y＝{y ₁ ,y ₂ }#(2)

where x represents the decision space and y represents the target space. In case of N prediction samples and M submodels, y ₁ Indicating the accuracy of the submodel, y ₂ Indicating diversity between submodels. F ⁱ Representing the actual value of the ith sample;

respectively representing the predicted values of the m-th model and the l-th model on the ith sample;

represents the average predicted value of all submodels at the ith sample.

Step 2: initializing the algorithm population, and generating a primary sub-model in the initialization process.

And 3, step 3: and generating a next generation of individuals by using a differential evolution algorithm of 2 different search modes according to the pareto dominant states of the selected individuals.

And 4, step 4: and iterating the population by using an angle and distance-based environment selection method.

And 5, step 5: and repeating the step 3 and the step 4 until the iteration number meets the requirement.

And 6, a step of: using an optimal selection strategy to remove individuals with poor population performance, wherein the formula is as follows:

wherein i is the number of objective functions and y is the objective function.

The generation and optimization of the sub-model are realized through the steps, and the specific implementation is as shown in fig. 1, wherein i represents the current iteration times, and G represents the preset total iteration times

Based on MOGRU nonlinear integration strategy:

step 1: and taking the training outputs of the load prediction submodel and the error prediction submodel as the training input of the GRU model.

Step 2: and taking the load actual value as the training output of the GRU model.

And 3, step 3: and optimizing GRU model parameters by using the improved MOEA/D-SAS, and training to obtain the MOGRU model.

When prediction is carried out, the outputs of the load prediction submodel and the error prediction submodel are used as the input of the MOGRU, and the obtained output is the final prediction result.

Through the steps, the integrated output of the plurality of GRU error prediction and load prediction submodels is realized. The workflow diagram of the MOGRU nonlinear integration strategy is shown in FIG. 2.

A multi-stage integrated prediction framework based on error factors and improved MOEA \ D is as follows:

the first stage is as follows: the CEEMDAN is used to decompose the original data into subsequences with different characteristic dimensions, such as IMF1, IMF2, \ 8230, IMFK. Each sub-sequence is then predicted using a parameter uniform GRU. At this point, we will get the initial predicted sequences Y1, Y2, \8230, YK. In order to maintain data integrity, historical information is extracted conveniently, and the obtained prediction sequence is combined with original data to obtain a new data set.

And a second stage: and carrying out supervised training on the plurality of MOGRU models by using the training set, predicting the training set and the test set after the models are converged, and obtaining a plurality of groups of training set predicted values and test set predicted values. And subtracting the actual value of the training set from the predicted value of the training set to obtain a plurality of groups of error sequences. And training the MOGRU by using the error sequence, and then carrying out error prediction on the test set to obtain error prediction values of a plurality of groups of test sets.

And a third stage: and integrating and outputting the load predicted values and the error predicted values of the plurality of groups of test sets by using a nonlinear integration method based on the MOGRU to obtain a final load predicted value.

A multi-stage integration flow chart based on error factors and improved MOEA \ D-SAS is shown in FIG. 3.

Claims (2)

1. The multi-stage integrated short-term load prediction based on the error factor and the improved MOEA \ D-SAS comprises the following steps:

the components of the CEEMDAN decomposition are predicted using GRUs and combined with the original data set to produce a new data set.

And generating a plurality of GRU load prediction and error prediction submodels through an improved MOEA/D-SAS optimization algorithm, and optimizing the submodels.

And (4) screening the sub-models for 2 times by using an optimal selection strategy to remove the sub-models with poor performance.

And integrating and outputting the load prediction submodel and the error prediction submodel by using a nonlinear integration method based on the MOGRU to obtain a final load prediction value.

The improved MOEA/D-SAS algorithm is characterized in that an environment selection method and a self-adaptive population generation strategy based on angles and distances are provided. The mathematical model is as follows:

(1) Angle and distance based environment selection

Wherein F (x) represents a mapping of an individual x in a target space, z ^* Is an ideal point, λ ^j Is the jth weight vector. Compared with the original MOEA \ D-SAS environment selection method, the environment selection method takes the distance factor between an individual and an ideal point into consideration. The situations of good distribution and poor convergence of the population are avoided. In practical engineering application, the improved environment selection method strengthens selection pressure and overcomes the defect of seed selectionUnder the conditions of small group and few iteration times, the selection of elite individuals is missed, and the quality of the generated model is improved.

(2) Adaptive population generation strategy

Wherein v is ⁱ Denotes a mutant individual, x ⁱ Indicates the current individual or individuals to be present,

representing 4 different individuals randomly selected by the current population. The traditional MOEA \ D-SAS algorithm is difficult to explore a decision space under the conditions of few populations and iteration times. The strategy uses DE mutation strategies of different search modes to explore space according to the pareto dominant state of the selected individual, and further improves the search capability of the algorithm.

2. The multi-stage integrated short-term load prediction method based on error factors and improved MOEA \ D-SAS algorithm as recited in claim 1, wherein: predicting the CEEMDAN decomposed components by adopting GRU, combining the components with the original data set to obtain a new data set, and fully mining data characteristics; the GRU learning rate and the number of hidden nodes in each layer are used as a decision space of an optimization algorithm, the accuracy and diversity of a GRU network are used as a target space of the optimization algorithm to optimize the submodel, and the accuracy of the submodel is improved while the diversity of the submodel is ensured; selecting the optimal selection strategy to eliminate the individual with problems in the optimization result, and providing the most appropriate sub-model for integrated output; a sub-model generation strategy based on improved MOEA \ D-SAS is used for establishing a load prediction model and an error prediction model, and the influence of data decomposition and model defects on a prediction result is reduced; and integrating and outputting the load prediction and error prediction submodels by using a nonlinear integration strategy based on the MOGRU, and further improving the prediction precision by considering error factors.

Images