Power load prediction method and device based on hybrid ELM algorithm, electronic equipment and storage medium

Abstract

The invention discloses a power load prediction method and device based on a hybrid ELM algorithm, electronic equipment and a storage medium, and the method achieves the prediction of a power load through an ELM prediction model (MT-ELM) mixed with MINRES and hidden layer traversal, can achieve the estimation of the number of optimal hidden layer nodes, and eliminates the overfitting. Therefore, the prediction level of the power load is effectively improved. The invention discloses a power load prediction method based on a hybrid ELM algorithm. The method comprises the following steps: S1, performing variation on input and output training sets to construct a training set suitable for an ELM prediction model; s2, calculating a hidden layer output matrix H of the ELM neural network through a calculation mode of a classic ELM model; and S3, traversing and estimating the number of optimal hidden points by using the input and output test sets, eliminating ELM over-fitting through a mixed MINRES method, learning through data over the years and generating an ELM mixed model, and predicting the future short-term power load condition through the generated model.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence computer algorithms and power systems, to the construction of algorithm models and the training and prediction of actual data, and is specifically used for power load prediction. Background Art

[0002] At present, the power supply system is facing the challenges of power supply security and green power transformation. Promoting the transformation of the power energy system to green and low-carbon has become a common goal of power energy development planning for countries around the world. In recent years, China has been promoting the scale and proportion of renewable energy installed capacity. With the continuous increase in the scale and proportion of new energy installed capacity, the original power system model is no longer applicable. Here, an ultra-high proportion of renewable energy power system is constructed. The focus is on achieving a balance between flexible resources and regulation needs. This is the development trend of the future power system. Power load forecasting is an important part of the safe and stable operation of the power system, and it is also an important part of the development planning and dispatching operation of the power grid.

[0003] Extreme Learning Machine (ELM) is a new neural network method in recent years. It is popular due to its fast training speed and high accuracy, and has a place in power load forecasting. However, the traditional ELM forecasting model is overfitting and the number of hidden nodes is usually much smaller than the input node value, which poses a new challenge to the output of the model. Summary of the invention

[0004] The technical problem to be solved by the present invention is that in view of the deficiencies in the above-mentioned prior art, the present invention discloses a power load forecasting method based on a hybrid ELM algorithm. The present invention predicts the power load by hybrid MINRES and an ELM forecasting model of hidden layer traversal (MT-ELM), which can estimate the optimal number of hidden layer nodes and eliminate overfitting, thereby effectively improving the prediction level of the power load.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] In a first aspect, the present invention discloses a method for predicting power load based on a hybrid ELM algorithm.

[0007] The following steps are included:

[0008] S1: Vary the input and output training sets to construct a training set suitable for the ELM prediction model;

[0009] S2: Calculate the hidden layer output matrix H of the ELM neural network using the calculation method of the classic ELM model;

[0010] S3: Use the input and output test sets to traverse and estimate the optimal number of hidden points, eliminate ELM overfitting through the hybrid MINRES method, learn and generate the ELM hybrid model through historical data, and use the generated model to predict future short-term power load conditions.

[0011] As an improvement of a power load forecasting method disclosed in the present invention,

[0012] In step S1:

[0013] If any N samples (x i , t i ), i = 1, 2, ..., N; here t i =[t i1 ,t i2 ,…t iN ] T ∈R m , x i =[x i1 ,x i2 ,…x iN ] T ∈R n ;x i is the input training sample, t i To output training samples; the model generates an output set through L neural network hidden nodes and activation function g(.):

[0014]

[0015] In formula (1), w j =[w j1 ,w j2 ,…,w jn ] T is the weight vector of the jth hidden node and the output layer, β i =[β i1 ,β i2 ,…β im ] T is the weight vector from the ith hidden node to the output node, and b i is the threshold of the i-th hidden node; w j ·x i w j With x i The inner product of; for L hidden nodes with activation function g(.), the zero error is approximately estimated as Σ i＝1 L =||o j -t j ||=0, here g(.) is infinitely differentiable;

[0016] β j、w j and b j Described as:

[0017]

[0018] For L formulas, it simplifies to:

[0019] Hβ=T (3)

[0020] here

[0021]

[0022]

[0023] In formula (3) and formula (4), H is the hidden layer output matrix of the neural network, and the i-th column of H also represents the output of the i-th hidden node of a set of input training sets;

[0024] Select L≤N.

[0025] As an improvement of a power load forecasting method disclosed in the present invention,

[0026] The classic ELM algorithm of step S2 is:

[0027] First, a set of input weight vectors w is randomly generated i and threshold vector b i , i = 1, 2, ... L; secondly, calculate the hidden layer output matrix of the neural network; and again calculate a set of output weight vectors β

[0028]

[0029] In formula (5), H ↑ is the Moore-Penrose generalized inverse matrix of the hidden layer output matrix H of the neural network;

[0030] in:

[0031]

[0032] As an improvement of a power load forecasting method disclosed in the present invention,

[0033] The variant of step S1 for inputting and outputting training sets is:

[0034] Given a power load data e=[e1,e2,…e N ] T ∈R k ; where e ijRepresents the power load data at time j-1 on the i-th day; construct input training samples through output data. Assuming that a data set of m days is used, a window of s days is used to predict the data of the p-th day, and construct input and output training sets:

[0035] make

[0036]

[0037]

[0038] In formula (7), each training set x i Represents the data of a window of size s days at all times; each training set t i Represents the power load data that occurs every day in a window of 1 day;

[0039] make

[0040]

[0041] In formula (8), the input test set x testi Represents the data at all times based on m+1 days and in a window of size s days; output test set t i Represents the daily power load data in the window of day m+s+1;

[0042] Correct the above formula (3):

[0043] H T Hβ=H T T

[0044]

[0045] As an improvement of a power load forecasting method disclosed in the present invention,

[0046] The process of traversing the optimal number of hidden nodes included in step S3 is as follows:

[0047] S301: Select the number of hidden layer nodes;

[0048] S302: Calculate the hidden layer output matrix;

[0049] S303: construct ELM model;

[0050] S304: constructing an input / output test matrix;

[0051] S305: Calculate output weight matrix;

[0052] S306: Calculate prediction results.

[0053] In a second aspect, the present invention also discloses a power load forecasting device based on a hybrid ELM algorithm.

[0054] Include:

[0055] Training set construction unit: used to transform the input and output training sets to construct a training set suitable for the ELM prediction model;

[0056] Computation unit: Calculate the hidden layer output matrix H of the ELM neural network using the calculation method of the classic ELM model;

[0057] Prediction unit: Use the input and output test sets to traverse and estimate the optimal number of hidden points, eliminate ELM overfitting through the hybrid MINRES method, learn and generate the ELM hybrid model through historical data, and use the generated model to predict the future short-term power load situation.

[0058] In a third aspect, the present invention further discloses an electronic device,

[0059] include:

[0060] one or more processors;

[0061] A memory, coupled to the processor, for storing one or more programs;

[0062] When the one or more programs are executed by the one or more processors, the one or more processors implement any one of the above-mentioned power load forecasting methods.

[0063] In a fourth aspect, the present invention further discloses a computer-readable storage medium having a computer program stored thereon.

[0064] When the computer program is executed by a processor, any one of the above-mentioned power load forecasting methods is implemented.

[0065] The beneficial effects of the present invention are:

[0066] The MT-ELM prediction model formed by the present invention through the method of mixing MINRES and traversing the optimal number of hidden layer nodes predicts the power load in the future period, which has good hidden measurement number screening characteristics and the suppression of overfitting effect, and significantly improves the classic ELM power load prediction model. Through the present invention, the ELM hybrid model can be quickly learned and generated through historical data, and the generated model can be used to predict the future short-term load situation, which can realize the estimation of the optimal number of hidden layer nodes and eliminate overfitting, thereby effectively improving the prediction level of power load. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a schematic diagram of the model structure of the extreme learning machine (ELM);

[0068] Figure 2 The figure is a flow chart of a specific implementation mode of the present invention. DETAILED DESCRIPTION

[0069] The specific implementation of the present invention is described below in conjunction with the accompanying drawings and embodiments:

[0070] As shown in the figure, it shows a specific embodiment of the present invention.

[0071] If any N samples (x i , t i ), i = 1, 2, ..., N. Here t i =[t i1 ,t i2 ,…t iN ] T ∈R m , x i =[x i1 ,x i2 ,…x iN ] T ∈R n .x i is the input training sample. The model generates an output set through L neural network hidden nodes and activation function g(.):

[0072]

[0073] Here j =[w j1 ,w j2 ,…,w jn ] T is the weight vector of the jth hidden node and the output layer, β j =[β j1 ,β j2 ,…β jm ] T is the weight vector from the jth hidden node to the output node, and b j is the threshold of the jth hidden node. j ·x i w j With x i For L hidden nodes with activation function g(.), the zero error can be approximately estimated as Σ i＝1 L =||o j -t j ||=0, where g(.) is infinitely differentiable (it can be a symbolic function, vector machine, sine function, cosine function, exponential function and non-regular function, etc.) exists;

[0074] βj 、w j and b j It can be described as:

[0075]

[0076] For L formulas, it can be simplified to:

[0077] Hβ=T (3)

[0078] here

[0079]

[0080] H is the hidden layer output matrix of the neural network, and the i-th column of H also represents the output of the i-th hidden node of a set of input training sets.

[0081] When L>N, the linear system has many zero-error solutions, but it may cause overfitting, so L≤N is usually selected.

[0082] The classic ELM algorithm is as follows:

[0083] First, a set of input weight vectors w is randomly generated i and threshold vector b i , i = 1, 2, ... L. Next, calculate the hidden layer output matrix of the neural network. Again calculate a set of output weight vectors β

[0084]

[0085] Here H ↑ is the Moore-Penrose generalized inverse matrix of the hidden layer output matrix H of the neural network.

[0086] in:

[0087]

[0088] The classic ELM model can only be applied to fitting and classification. The following is a variation of the input and output training sets to achieve the purpose of prediction.

[0089] Given a power load data e=[e1,e2,…e N ] T ∈R k . Where e ij Represents the power load data at time j-1 on the i-th day. The input training samples are constructed through the output data. Assuming that the m-day data set is used, a window of s days (the default step size is 1) is used to predict the data of the p-th day, and the input and output training sets are constructed.

[0090]

[0091] Among them, each training set x i Represents the data of a window of size s days at all times. Each training set t i Represents the power load data that occurs every day in a window of 1 day.

[0092] make

[0093]

[0094] Among them, the input test set x testi Represents the data at all times based on m+1 days and in a window of size s days. Output test set t i Represents the daily power load data in the window of day m+s+1.

[0095] Through the above ELM prediction model, the output weight matrix can be obtained to build a complete ELM prediction model. However, usually the number of hidden nodes It is a manual estimate. Here, the optimal number of hidden nodes can be obtained by comparing the output of the test set with the original data in a traversal manner. Here, due to the arbitrariness of the input weight matrix and bias matrix, the optimal number of hidden nodes is is a random variable. In the present invention, the average value of several traversals is taken to estimate the optimal number of hidden nodes.

[0096] In the actual process, the optimal number of hidden nodes may exceed the number of training sets, resulting in overfitting of the output results. In order to eliminate the overfitting phenomenon, formula (3) is modified here:

[0097] H T Hβ=H T T

[0098]

[0099] The above embodiment forms an MT-ELM prediction model by mixing MINRES and traversing the optimal number of hidden layer nodes to predict the power load in the future period. It has good hidden measurement number screening characteristics and the suppression of overfitting effects, which significantly improves the classic ELM power load prediction model.

[0100] The model structure of the extreme learning machine (ELM) is as follows Figure 1 As shown, it includes input layer, hidden layer and output layer, and the layers are closely connected in sequence.

[0101] The hybrid MINRES and hidden layer traversal ELM prediction method of the present invention proposes methods from three aspects. The specific process is as follows: Figure 2 shown.

[0102] First, the input and output training sets are modified to construct a training set suitable for the ELM prediction model, and the hidden layer output matrix H of the ELM neural network is calculated by the calculation method of the classic ELM model. Since the calculation of the classic ELM model is relatively simple, the present invention uses the input and output test sets to traverse and estimate the optimal number of hidden points, and eliminates ELM overfitting through the hybrid MINRES method, and improves the prediction quality of the data. The method of the present invention can quickly learn and generate an ELM hybrid model through historical data, and predict the future short-term load situation through the generated model.

[0103] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments. Various changes can be made within the knowledge scope of ordinary technicians in the field without departing from the purpose of the present invention. These changes involve related technologies well known to those skilled in the art, which all fall within the scope of protection of the patent of the present invention.

[0104] Many other changes and modifications may be made without departing from the concept and scope of the present invention.It should be understood that the present invention is not limited to the specific embodiments, and the scope of the present invention is defined by the appended claims.

Claims

1. Power load forecasting method based on hybrid ELM algorithm, It is characterized in that The following steps are included: S1: Vary the input and output training sets to construct a training set suitable for the ELM prediction model; S2: Calculate the hidden layer output matrix H of the ELM neural network using the calculation method of the classic ELM model; S3: Use the input and output test sets to traverse and estimate the optimal number of hidden points, eliminate ELM overfitting through the hybrid MINRES method, learn and generate the ELM hybrid model through historical data, and use the generated model to predict future short-term power load conditions.

2. The power load forecasting method based on the hybrid ELM algorithm according to claim 1, Features: In step S1: If any N samples (x i , t i ), i = 1, 2, ..., N; here t i =[t i1 ,t i2 ,…t iN ] T ∈R m , x i =[x i1 ,x i2 ,…x iN ] T ∈R n ;x i is the input training sample, t i To output training samples; the model generates an output set through L neural network hidden nodes and activation function g(.): In formula (1), w j =[w j1 ,w j2 ,…,w jn ] T is the weight vector of the jth hidden node and the output layer, β i =[β i1 ,β i2 ,…β im ] T is the weight vector from the ith hidden node to the output node, and b i is the threshold of the i-th hidden node; w j ·x i w j With x i The inner product of; for L hidden nodes with activation function g(.), the zero error is approximately estimated as Σ i＝1 L =||o j -t j ||=0, here g(.) is infinitely differentiable; β j 、w j and b j Described as: For L formulas, it simplifies to: Hβ=T (3) here In formula (3) and formula (4), H is the hidden layer output matrix of the neural network, and the i-th column of H also represents the output of the i-th hidden node of a set of input training sets; Select L≤N.

3. The power load forecasting method based on the ELM algorithm according to claim 1, Features: The classic ELM algorithm of step S2 is: First, a set of input weight vectors w is randomly generated i and threshold vector b i , i = 1, 2, ... L; secondly, calculate the hidden layer output matrix of the neural network; and again calculate a set of output weight vectors β In formula (5), H ↑ is the Moore-Penrose generalized inverse matrix of the hidden layer output matrix H of the neural network; in:

4. The power load forecasting method based on the hybrid ELM algorithm as claimed in claim 2, Features: The variant of step S1 for inputting and outputting training sets is: Given a power load data e=[e1,e2,…e N ] T ∈R k ; where e ij Represents the power load data at time j-1 on the i-th day; construct input training samples through output data. Assuming that a data set of m days is used, a window of s days is used to predict the data of the p-th day, and construct input and output training sets: make In formula (7), each training set x i Represents the data of a window of size s days at all times; each training set t i Represents the power load data that occurs every day in a window of 1 day; make In formula (8), the input test set x testi Represents the data at all times based on m+1 days and in a window of size s days; output test set t i Represents the daily power load data in the window of day m+s+1; Correct the above formula (3):

5. The power load forecasting device based on the hybrid ELM algorithm according to claim 1, It is characterized in that The process of traversing the optimal number of hidden nodes included in step S3 is as follows: S301: Select the number of hidden layer nodes; S302: Calculate the hidden layer output matrix; S303: construct ELM model; S304: constructing an input / output test matrix; S305: Calculate output weight matrix; S306: Calculate prediction results.

6. Power load forecasting device based on hybrid ELM algorithm, It is characterized in that Include: Training set construction unit: used to transform the input and output training sets to construct a training set suitable for the ELM prediction model; Computation unit: Calculate the hidden layer output matrix H of the ELM neural network using the calculation method of the classic ELM model; Prediction unit: Use the input and output test sets to traverse and estimate the optimal number of hidden points, eliminate ELM overfitting through the hybrid MINRES method, learn and generate the ELM hybrid model through historical data, and use the generated model to predict the future short-term power load situation.

7. An electronic device, It is characterized in that include: one or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the power load forecasting method as described in any one of claims 1 to 4.

8. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the power load forecasting method according to any one of claims 1 to 4 is implemented.

Images