CN117543537A

CN117543537A - Agent electricity purchasing user electric quantity prediction method, device and storage medium

Info

Publication number: CN117543537A
Application number: CN202311328853.3A
Authority: CN
Inventors: 魏立勇; 丁一; 周颖; 王永利; 王恩; 董焕然; 庞超; 邱敏; 霍现旭; 刘念; 赵伟博; 陈亮; 贺春; 白雪峰; 张剑; 赵晨阳; 李熠; 吴磊; 陈宋宋
Original assignee: State Grid Corp of China SGCC; China Electric Power Research Institute Co Ltd CEPRI; State Grid Tianjin Electric Power Co Ltd; North China Electric Power University; Electric Power Research Institute of State Grid Tianjin Electric Power Co Ltd
Current assignee: State Grid Corp of China SGCC; China Electric Power Research Institute Co Ltd CEPRI; State Grid Tianjin Electric Power Co Ltd; North China Electric Power University; Electric Power Research Institute of State Grid Tianjin Electric Power Co Ltd
Priority date: 2023-10-15
Filing date: 2023-10-15
Publication date: 2024-02-09

Abstract

The invention relates to a proxy electricity purchasing user electricity quantity prediction method based on an echo state machine and error checking, which comprises the following steps: constructing a novel reserve pool according to the historical electricity consumption and time sequence data of the electricity consumption influence factors; constructing a multi-reservoir related vector echo state machine model based on the constructed novel reservoir; based on the constructed multi-reservoir related vector echo state machine model, an error check-based multi-reservoir related vector echo state machine model is established, historical electricity consumption data with different granularities and influence factor data of electricity consumption are input in a matrix form, and finally an electricity quantity prediction result is output. The method and the device can effectively improve the accuracy of the electricity consumption prediction of the agent electricity purchasing user.

Description

Agent electricity purchasing user electric quantity prediction method, device and storage medium

Technical Field

The invention belongs to the technical field of user electricity quantity prediction, relates to a method and a device for predicting the electricity quantity of a proxy electricity purchase user, and particularly relates to a method and a device for predicting the electricity quantity of a proxy electricity purchase user based on an echo state machine and error checking.

Background

The neural network is composed of an input layer, a hidden layer and an output layer, and can be divided into a surface level neural network and a deep layer neural network according to the number of layers of the hidden layer. A common shallow neural network is a BP neural network. For example, the dimension of the input space is reduced by adopting a kernel principal component analysis method, and then the BP neural network is subjected to parameter adjustment by using a particle swarm optimization algorithm, so that the prediction accuracy is effectively improved. Common deep neural networks are deep belief networks (DeepBeliefNetworks, DBN), extreme learning machines (Extreme Learning Machine, ELM), and Long Short-Term Memory (LSTM) neural networks. The memory enhancement network (Memory Augmented Networks, MAN) is a popular deep neural network method in the field of time sequence prediction, replaces the traditional long short-term memory (LSTM) neural network, and effectively solves the long-term dependence challenge by using long-term memory cells to replace the traditional hidden layer neurons.

Along with the continuous increase of the quantity and variety of energy data provided by building automation devices, intelligent electric meters and other resources, the noise of power data and the excessive simplicity of the super-parameter value selection of the LSTM neural network are not capable of generating satisfactory prediction results. Moreover, the neural network is easy to excessively fit on the training set, so that the new input data is not good, meanwhile, a plurality of related factors such as weather, seasons, holidays and the like need to be considered in the prediction model, but the correlation between the data is not easy to determine, and the inaccuracy of the prediction result can be caused.

In addition, in the electric power device, along with the development and improvement of the sensors of each department, a large amount of data of different types are collected, and in order to reduce the occurrence of a phenomenon that the calculation cost is greatly increased due to the input of irrelevant data, the redundant data needs to be reduced in dimension before the prediction of the electricity consumption column. The data dimension reduction is to reduce the data dimension on the premise of completely expressing the original data characteristics, further improve the quality of the input data of the prediction model and reduce the operation time. The mainstream dimension reduction algorithm at the present stage is Principal Component Analysis (PCA), kernel principal component analysis (Kernel PCA, KPCA), elastic Net (EN), and the like.

While a combined predictive model based on data dimension reduction is advantageous in some aspects, in order to reduce the data dimension during the data dimension reduction, a portion of the information of the original data must be lost. This may lead to loss of key features, affecting prediction accuracy. While the dimension reduction process requires selection of the remaining important features, determining which features are most significant for the agent's business electricity consumption prediction is not a simple task and erroneously selecting features may result in bias or inefficiency in the prediction model. And because the dimension reduction process may cause variance reduction of the data set, prediction capability of the model is limited, and in particular, in the case of nonlinear relationship, the complex relationship between data is likely to be not captured well by the dimension reduction of the data. The combined model is composed of a plurality of sub-models, each sub-model possibly has errors, when the prediction results of the sub-models are overlapped, the errors are overlapped, so that the inaccuracy of the whole prediction result is caused, and particularly when a certain sub-model has a larger error, the performance of the whole combined model is obviously affected.

Therefore, how to effectively improve the accuracy of the electricity consumption prediction of the agent electricity purchasing user is a technical problem to be solved by the technicians in the field.

No prior art patent document, which is the same as or similar to the present invention, was found after searching.

Disclosure of Invention

The invention aims to overcome the defects of the prior art, and provides a method and a device for predicting the electric quantity of a proxy electricity purchasing user based on an echo state machine and error checking, which can effectively improve the accuracy of the electric quantity prediction of the proxy electricity purchasing user.

The invention solves the practical problems by adopting the following technical scheme:

the agent electricity purchasing user electricity quantity prediction method based on the echo state machine and error check comprises the following steps:

constructing a novel reserve pool according to the historical electricity consumption and time sequence data of the electricity consumption influence factors;

constructing a multi-reservoir related vector echo state machine model based on the constructed novel reservoir;

based on the constructed multi-reservoir related vector echo state machine model, an error check-based multi-reservoir related vector echo state machine model is established, historical electricity consumption data with different granularities and influence factor data of electricity consumption are input in a matrix form, and finally an electricity quantity prediction result is output.

And the specific steps of constructing the novel reserve pool according to the historical electricity consumption and the time series data of the electricity consumption influence factors comprise:

Creating a pool comprising a plurality of neurons, each neuron representing a state node, selecting an appropriate number of state nodes as the base unit of the pool;

determining the dimension of a weight matrix according to the number of state nodes in the reserve pool, and randomly distributing input weight and echo weight for each state node;

the sparse connection matrix is used for representing internal connection weight in the echo state machine of the reserve pool, a sparse connection matrix is created by adjusting sparsity, elements of the matrix represent connection strength, a random number generator is used for generating the connection matrix, and whether connection exists or not is determined according to the sparsity at each time step;

at each time point, updating the state of the reserve tank according to the current state and input of the reserve tank, and assuming that N state nodes exist in the reserve tank, the current value of the state nodes is x (t), wherein t represents the current time step;

the process of updating a state node may be described as follows:

x(t)＝(1-a)x(t-1)+a×tanh[W _in u(t)+Wx(t-1)+W _back y(t-1)] (1)

in the above expression, it is assumed that x (t) represents the reservoir status value at the present time, x (t-1) represents the reservoir status value at the previous time, u (t) represents the input value at the present time, y (t-1)]The output value at the previous time is indicated, and a is the update rate. In this arrangement, W _in Represents the weight matrix of the input layer, W represents the weight matrix connected with the interior of the storage pool, W _back Representing a connection weight matrix of the output layer to the reserve tank; tanh () is an activation function that converts a linearly combined input into a nonlinear output, and a hyperbolic tangent function is used in the present invention.

Determining input and output modes according to the electricity consumption prediction requirement of the agent electricity purchasing user;

selecting the number of input nodes and the number of output nodes, and setting according to the dimension of input data and the output result;

subjecting the pool to idle operation for a period of time, i.e. discarding state variables at time t-1 before;

starting from time t, the pool will keep and update state variables and generate a new matrix to improve the initial state echo problem:

X＝[x(t),x(t+1),...,x(t+N)] ^T (2)

the output after pool mapping is:

y(X)＝W _out X (3)

in which W is _out Is an output weight matrix.

Training and adjusting the reserve pool by using known input data and last-step output data, and readjusting parameters such as connection weight, bias item and the like;

and evaluating the performance of the trained reserve pool by using the verification data set, performing necessary tuning, improving the accuracy and generalization capability of the echo state machine by adjusting parameters such as the number of state nodes, the connection strength, the input and output configuration and the like, and finally completing the establishment of the reserve pool.

Moreover, the specific steps of constructing the multi-reservoir correlation vector echo state machine model based on the constructed novel reservoir comprise the following steps:

for each reservoir, the state node value inside the reservoir is updated by using a state update equation, and the state update of the whole reservoir is processed in one operation by adopting a vectorization mode to generate a plurality of different reservoirs, X ₁ ,X ₂ ,...,X _n And then weighted combining to achieve the update.

The outputs of the plurality of reservoirs are connected together to form a final output layer, and the outputs of the plurality of reservoirs are integrated by a simple connection operation.

Wherein, the double reservoir is expressed as:

X＝λX ₁ +(1-λ)X ₂ (4)

in the above formula, λ represents the weight coefficient of each reservoir, and these parameters are adjusted by the particle swarm optimization algorithm to obtain the optimal value. Thereby forming a multi-reservoir combination.

The multi-reservoir combination is applied to a correlation vector machine, and a multi-reservoir correlation vector echo state machine MrRVESM model is obtained by replacing a kernel function in the vector machine;

the multi-reservoir correlation vector echo state machine model outputs are:

y＝ω ^T X+ε (5)

in the equation, epsilon represents a noise term introduced by the model, and omega represents an output weight matrix;

if there are N sample data, it is assumed that the ith column and its corresponding output weight are respectively denoted as X _i And omega _i The output vector is the predicted value of the electricity consumption of the agent electricity purchasing user, and can be expressed in the following way:

assume that a mean value of 0 and a variance of sigma are defined ² If the gaussian noise of (c) is epsilon, then the whole sample also satisfies the gaussian distribution, and the likelihood function of the sample can be expressed as:

in the formula, … ² The square value is taken after matrix norm operation is represented.

The weight matrix ω is added with terms that obey a 0-mean gaussian prior distribution. The specific expression is as follows:

wherein: omega= (omega) ₀ ,ω ₁ ,…,ω _i ,...,ω _N ) ^T The method comprises the steps of carrying out a first treatment on the surface of the Alpha is variance, alpha= (alpha) ₀ ,α ₁ ,...,α _i ,...,α _N ) ^T The method comprises the steps of carrying out a first treatment on the surface of the p (ω|α) represents the conditional probability of ω;is omega _i Is a normal distribution density function of (c).

By substituting the formula (8) into the formula (7) and converting into a definite integral form by using the bayesian theorem and simplifying, the posterior distribution of ω can be obtained, and the specific expression is as follows:

p(ω|y,α,σ ² )＝N(ω|μ,∑) (9)

in the equation, μ and Σ represent the mean and variance of the posterior distribution of ω, respectively, and their calculation formulas are as follows:

∑＝(σ ^-2 X ^T X+α) ^-1 (10)

μ＝σ ^-2 ∑X ^T y (11)

by solving for alpha and sigma ² The two super parameters can be used for obtaining weight distribution, and the partial derivative calculation is carried out on the two super parameters, and the partial derivative calculation is made to be zero, so that an iteration equation is obtained and is used for solving the update of the weight, and the expression is as follows:

wherein: gamma ray _i ＝1-α _i ∑ _i,i ，Σ _i,i Represents the diagonal elements of the sigma matrix, and mu _i The value obtained by continually iterating through the update (10).

Training the reserve pool by using marked training data, and improving the performance by adjusting the connection weight and other parameters according to the predicted power consumption requirement of the agent electricity purchasing user;

after training is completed, the data to be tested can be input, and a reserve pool matrix X is generated _* Prediction is carried out, and a prediction result y is finally output _* Can be expressed as:

y _* ＝μ ^T X _* (14)

and the specific steps of establishing the multi-reservoir related vector echo state machine model based on error check, inputting historical electricity consumption data and influence factor data of electricity consumption of different granularities in a matrix form, and finally outputting an electricity prediction result comprise:

the dimension and the format of input data are determined, the input data comprise information related to agent electricity purchasing business such as historical electricity purchasing data, market price data, generating capacity data and the like, normalization processing and daily type separation are carried out on the data, and a weight matrix is initialized.

During the training of the MrRVESM model, a time point t is selected after the end of the pool idling _i Obtaining training errors of each time point, and generating a brand-new error sequence; then, each reserve pool is weighted and combined, and the weight is optimized; finally, temperature and precipitation meteorological data which are obviously related with the electricity consumption prediction result are selected, a unique influence factor sequence is constructed, and an electricity preliminary prediction value y is obtained _t 。

And extracting the electricity consumption influence factors according to the electricity consumption characteristics of the electricity consumption of the agent electricity purchasing user. Combining the training error sequence with the influencing factor sequence to generate a brand-new training error set.

The new training error set is trained by using a correlation vector machine (RVM), and after the training is completed, the error at the current time point can be predicted to obtain an error compensation value, so that a corresponding new error prediction result is obtained.

According to the short-term electricity consumption prediction requirement of the agent electricity purchasing user, the performance is improved by adjusting the connection weight and other parameters; training the model by using marked training data, and fusing an error prediction result with an output value of the MrRVESM model by using an error checking algorithm to construct a multi-reservoir related vector echo state machine EC-RVMESM based on error result checking; and by checking and correcting the output result, the influence of errors is reduced. And further finishing checking the electric quantity prediction result of the agent electricity purchasing user, and obtaining a final checked prediction result y.

An agent electricity purchase user electricity quantity prediction device based on an echo state machine and error checking comprises:

the novel reserve pool construction module is used for constructing a novel reserve pool according to the historical electricity consumption and the time sequence data of the electricity consumption influence factors;

The multi-reservoir related vector echo state machine model building module builds a multi-reservoir related vector echo state machine model based on the built novel reservoir;

the multi-reservoir related vector echo state machine model building module based on error check builds a multi-reservoir related vector echo state machine model based on error check, inputs historical electricity consumption data and influence factor data of electricity consumption of different granularities in a matrix form, and finally outputs an electricity prediction result.

Moreover, the novel reserve pool construction module further comprises:

a state node definition module in the pool:

a random initialization state node weight module:

determining the dimension of a weight matrix according to the number of state nodes in the reserve pool; for each state node, randomly assigning an input weight and an echo weight;

the short-time memory creation module of the reserve pool is realized by the sparse connection matrix:

Dynamically updating the pool state module:

the process of updating the state node is described as follows:

x(t)＝(1-a)x(t-1)+a×tanh[W _in u(t)+Wx(t-1)+W _back y(t-1)] (1)

in the above expression, it is assumed that x (t) represents the reservoir status value at the present time, x (t-1) represents the reservoir status value at the previous time, u (t) represents the input value at the present time, y (t-1)]An output value indicating the previous time, a being the update rate; in this arrangement, W _in Represents the weight matrix of the input layer, W represents the weight matrix connected with the interior of the storage pool, W _back Representing a connection weight matrix of the output layer to the reserve tank; tanh () is an activation function;

input and output modules for setting a reserve pool:

X＝[x(t),x(t+1),...,x(t+N)] ^T (2)

The output after pool mapping is:

y(X)＝W _out X (3)

in which W is _out The weight matrix is output;

repeating training and adjusting the reserve pool module:

and (3) after verification and optimization, completing a reserve pool establishment module:

and evaluating the performance of the trained reserve pool by using the verification data set, performing necessary tuning, and improving the accuracy and generalization capability of the echo state machine by adjusting the number of state nodes, the connection strength and the input and output configuration parameters to finally finish the establishment of the reserve pool.

Moreover, the multi-pool correlation vector echo state machine model building module further comprises:

a single pool construction module for updating the state node value in each pool by using a state update equation, processing the state update of the whole pool in one operation by adopting a vectorization mode to generate a plurality of different pools, X ₁ ,X ₂ ,…,X _n Then, weighted combination is carried out to realize updating;

connecting a plurality of storage pool connecting modules, connecting the outputs of the storage pools together to form a final output layer, and integrating the outputs of the storage pools through simple connecting operation;

Wherein, the double reservoir is expressed as:

X＝λX ₁ +(1-λ)X ₂ (4)

in the above formula, λ represents the weight coefficient of each pool, and these parameters are adjusted by the particle swarm optimization algorithm to obtain an optimal value; thereby forming a multi-reservoir combination;

the multi-reservoir related vector echo state machine forming module is applied to a related vector machine by adopting multi-reservoir combination, and a multi-reservoir related vector echo state machine MrRVESM model is obtained by replacing a kernel function in the vector machine;

the multi-reservoir correlation vector echo state machine model outputs are:

y＝ω ^T X+ε (5)

if there are N sample data, it is assumed that the ith column and its corresponding output weight are respectively denoted as X _i And omega _i The output vector is the predicted value of the electricity consumption of the agent electricity purchasing user, and the electricity consumption can be calculated by the following methodThe expression is carried out in the following way:

in the formula, … ² Representing matrix norm operation and then taking a square value;

adding a term obeying 0-mean Gaussian prior distribution to the weight matrix omega; the specific expression is as follows:

Wherein: omega= (omega) ₀ ,ω ₁ ,…,ω _i ,...,ω _N ) ^T The method comprises the steps of carrying out a first treatment on the surface of the Alpha is variance, alpha= (alpha) ₀ ,α ₁ ,...,α _i ,...,α _N ) ^T The method comprises the steps of carrying out a first treatment on the surface of the p (ω|α) represents the conditional probability of ω;is omega _i Is a normal distribution density function of (2);

p(ω|y,α,σ ² )＝N(ω|μ,∑) (9)

∑＝(σ ^-2 X ^T X+α) ^-1 (10)

μ＝σ ^-2 ∑X ^T y (11)

wherein: gamma ray _i ＝1-α _i ∑ _i,i ，∑ _i,i Represents the diagonal elements of the sigma matrix, and mu _i Then the value obtained by continually iterating through the update (10);

y _* ＝μ ^T X _* (14)。

moreover, the multi-reservoir related vector echo state machine model building module based on error checking further comprises:

The data input and normalization processing module is used for determining the dimension and format of input data, wherein the input data comprises information related to agent power purchase business such as historical power purchase data, market price data, generating capacity data and the like, carrying out normalization processing and daily type separation on the data, and initializing a weight matrix;

training module for MrRVESM model, at MDuring rRVESM model training, a time point t is selected after the idling of the reserve pool is finished _i Obtaining training errors of each time point, and generating a brand-new error sequence; then, each reserve pool is weighted and combined, and the weight is optimized; finally, weather data such as temperature, precipitation and the like which are obviously related with the electricity consumption prediction result are selected, and a unique influence factor sequence is constructed to obtain an electricity preliminary prediction value y _t ；

A new error training set module for fusing the error sequence and the influence factors is constructed, and the influence factors of the electricity consumption are extracted according to the electricity consumption characteristics of the electricity consumption of the agent electricity purchasing user; combining the training error sequence with the influence factor sequence to generate a brand-new training error set;

training and obtaining a new error sequence result module, training the new training error set by using a correlation vector machine model RVM, and predicting the error of the current time point after the training is completed to obtain an error compensation value, thereby obtaining a corresponding new error prediction result;

The preliminary electricity consumption prediction value checking module is used for improving the performance by adjusting the connection weight and other parameters according to the short-term electricity consumption prediction requirement of the agent electricity purchasing user; training the model by using marked training data, and fusing an error prediction result with an output value of the MrRVESM model by using an error checking algorithm to construct a multi-reservoir related vector echo state machine EC-RVMESM based on error result checking; the output result is checked and corrected, so that the influence of errors is reduced; and further finishing checking the electric quantity prediction result of the agent electricity purchasing user, and obtaining a final checked prediction result y.

A computer device comprising a memory and a processor, the memory storing a computer program which, when executed by the processor, causes the processor to perform the steps of the power prediction method.

A computer readable storage medium storing a computer program which, when executed by a processor, causes the processor to perform the steps of the power prediction method.

The invention has the advantages and beneficial effects that:

1. the invention provides a method and a device for predicting the electric quantity of a proxy electricity purchasing user based on an echo state machine and error checking, which are used for simplifying the process of kernel function selection and parameter optimization by using a brand-new support proxy electricity purchasing service model named EC-MrRVESM (error compensation-multi-reservoir relevance vector echostate machine) and mapping input data from a lower dimension to a high dimension space by using a plurality of preparation pool related vector echo state machines and error correction. In addition, the multi-reservoir has higher sparsity and more stable state than the kernel function, and the model shows more excellent nonlinear analysis capability and can more accurately read the dynamic characteristics of the device relative to the single reservoir. Meanwhile, the model also adopts a sparse Bayesian regression technology to obtain an output weight solution with sparsity, so that training time is shortened and calculation speed is improved. Finally, an independent error sequence is extracted from the randomly initialized weight and training errors, a plurality of influence factors are combined as input data, and the model is used for training, so that an error check predicted value is obtained, the original error is corrected, the predicted result is corrected, and the accuracy of the power consumption prediction of the agent electricity purchasing user is effectively improved.

2. The invention utilizes a plurality of preparation pool related vector echo state machines and error correction to map input data from a lower dimension to a high dimension space so as to replace the traditional kernel function, simplify the process of kernel function selection and parameter optimization, and have higher sparsity and more stable state. Meanwhile, the model also adopts a sparse Bayesian regression technology, so that training time is shortened, and calculation speed is improved. The independent error sequence is extracted from the randomly initialized weight and training error, and the training is performed by using a related vector regression model, so that an error check measured value is obtained by the model, a prediction result is corrected, and the accuracy of the power consumption prediction of the agent electricity purchasing user is effectively improved.

Drawings

FIG. 1 is a schematic diagram of a reserve tank structure of the present invention;

FIG. 2 is a block diagram of an MrRVESM of the present invention;

FIG. 3 is a flow chart of the EC-MrRVESM model process of the present invention.

Detailed Description

Embodiments of the invention are described in further detail below with reference to the attached drawing figures:

In this embodiment, as shown in fig. 1, the novel pool constructed in the step 1 adopts a structure distinct from that of a conventional neural network. The reserve pool is a random and scarce recursive structure with excellent short-term memory. Compared with the traditional neural network, the reservoir carries a large number of neurons, generally between 200 and 2000, and the connection weight matrix between the neurons keeps random and rare properties, and the sparsity is kept between 1% and 5%.

The specific steps of constructing the novel reserve pool comprise:

defining state nodes in the pool:

in this embodiment, the state nodes may be discrete nodes or continuous nodes, such as discrete or real values, each representing the state of the pool at a particular time step. The values of these status nodes may be expressed as degrees of activation or outputs of neurons. The value of the state node is updated over time, calculated according to the dynamic update formula of the reservoir. These values can be used as an intermediate representation of the pool echo state machine for prediction by the subsequent output layer.

Randomly initializing state node weights:

the dimension of the weight matrix is determined from the number of state nodes in the pool, and assuming that the number of state nodes is N, the dimension of the weight matrix will be N x N. For each state node, an input weight and an echo weight are randomly assigned.

In this embodiment, the input weight determines the influence of the input data on the state node, and the echo weight determines the connection strength between the state node and itself. The weight matrix may be scaled as needed to ensure that the range of weights is appropriate for the activation function of the reservoir and the range of the input signal.

Creating a sparse connection matrix to realize short-term memory of the reserve pool:

the sparse connection matrix is used for representing internal connection weight in the echo state machine of the reserve pool, a sparse connection matrix is created by adjusting the sparsity, the elements of the matrix represent connection strength, a random number generator is used for generating the connection matrix, and whether connection exists or not is determined according to the sparsity at each time step.

Dynamically updating pool states:

at each time point, the state of the pool is updated according to the current state and input of the pool, and the current value of the state node is x (t) assuming that there are N state nodes in the pool, wherein t represents the current time step. The process of updating a state node may be described as follows:

x(t)＝(1-a)x(t-1)+a×tanh[W _in u(t)+Wx(t-1)+W _back y(t-1)] (1)

In the above expression, it is assumed that x (t) represents the reservoir status value at the present time, x (t-1) represents the reservoir status value at the previous time, u (t) represents the input value at the present time, y (t-1)]The output value at the previous time is indicated, and a is the update rate. In this arrangement, W _in Represents the weight matrix of the input layer, W represents the weight matrix connected with the interior of the storage pool, W _back Representing a matrix of connection weights for the output layer to the pool. tanh () is an activation function that converts a linearly combined input into a nonlinear output, and a hyperbolic tangent function is used in the present invention.

Setting input and output of a reserve pool:

and determining input and output modes according to the predicted power consumption requirement of the agent electricity purchasing user. Selecting the number of input nodes and the number of output nodes, and setting according to the dimension of input data and the output result;

considering that the initial state attribute is poor, the echo problem of the reserve tank is relatively serious, so that the reserve tank can be subjected to idle operation in a period of time, namely, the state variable at the moment t-1 before giving up is given up;

X＝[x(t),x(t+1),...,x(t+N)] ^T (2)

the output after pool mapping is:

y(X)＝W _out X (3)

In which W is _out Is an output weight matrix.

Repeated training and adjustment of the reservoir:

training and adjusting the reserve tank by using known input data and last-step output data, and enabling the reserve tank to accurately predict an output result by readjusting parameters such as connection weight, bias items and the like.

And (3) after verification and tuning, establishing a reserve pool:

The reservoir structure is shown in fig. 1.

in this embodiment, a single reservoir cannot fully analyze the characteristics of each variable when mapping a multi-variable device. Since the connection weight matrix of the pool has highly random and sparse properties, considerable prediction errors may result in cases where the variable regularity is low. Thus, a multi-reservoir structure is introduced and an independent weight matrix is generated for each reservoir. After all the storage pools are updated, information fusion is carried out by introducing a relevant mechanism, so that the prediction accuracy can be improved better, and the storage pools are combined according to weights to form a brand new storage pool which shows stronger nonlinear analysis capability. Such a design promotes modeling ability of the model for the multi-variable device and reduces the risk of prediction error.

The specific steps of constructing the multi-reservoir related vector echo state machine include:

constructing a single reserve pool:

A plurality of reservoirs are connected:

the outputs of the multiple reservoirs are connected together to form a final output layer, and the outputs of the multiple reservoirs are integrated by a simple connecting operation, such as splicing or series connection.

Taking a double reservoir as an example, it can be expressed as:

X＝λX ₁ +(1-λ)X ₂ (4)

Forming a multi-pool correlation vector echo state machine:

the multi-reservoir combination is applied to a correlation vector machine, and a prediction model can be enhanced to process large-batch and dynamic data by replacing a kernel function in the vector machine, so that a multi-reservoir correlation vector echo state machine MrRVESM model is obtained, and the MrRVESM structure is shown in figure 2.

The multi-reservoir correlation vector echo state machine model outputs are:

y＝ω ^T X+ε (5)

in the above equation, ε represents the noise term introduced by the model and ω represents the output weight matrix.

in the above, I ² The square value is taken after matrix norm operation is represented.

To avoid the problem of too many support vectors similar to those in a support vector machine, a term obeying a 0-mean gaussian prior distribution is added to the weight matrix ω. The specific expression is as follows:

wherein: omega= (omega) ₀ ,ω ₁ ,...,ω _i ,...,ω _N ) ^T The method comprises the steps of carrying out a first treatment on the surface of the Alpha is variance, alpha= (alpha) ₀ ,α ₁ ,...,α _i ,...,α _N ) ^T The method comprises the steps of carrying out a first treatment on the surface of the p (ω|α) represents the conditional probability of ω; n (omega) _i |0,α _i ^-1 ) Is omega _i Is a normal distribution density function of (c).

p(ω|y,α,σ ² )＝N(ω|μ,∑) (9)

∑＝(σ ^-2 X ^T X+α) ^-1 (10)

μ＝σ ^-2 ∑X ^T y (12)

wherein: gamma ray _i ＝1-α _i ∑ _i,i ，∑ _i,i Represents the diagonal elements of the sigma matrix, and mu _i The value obtained by continually iterating through the update (10).

Training the reserve pool by using marked training data, and improving the performance by adjusting connection weight and other parameters according to the predicted power consumption requirement of the agent electricity purchasing user;

y _* ＝μ ^T X _* (14)

establishing an ECMrRVESM model of a multi-reservoir related vector echo state machine based on error check, and outputting an electric quantity prediction result;

in this embodiment, the time sequence of the electricity consumption of the agent electricity purchasing user generally shows a periodic variation rule, so that similar results can be obtained in prediction. However, in handling multivariable nonlinear devices, other influencing factors (e.g. climate conditions) often have uncertainties, and the weight matrix in the model reservoir also has some randomness, which will lead to deviations in the predicted outcome. Therefore, the model prediction result is checked by introducing an error check link so as to correct the deviation of the prediction result. And finally, forming an ECMrRVESM model which realizes an error checking function on the basis of a multi-reservoir correlation vector echo state machine. And the checked error prediction result is fused with the prediction value output by the MrRVESM model, so that the power consumption prediction result of the MrRVESM model is checked, and the accuracy of the power consumption prediction result of the agent electricity purchasing user group is improved.

As shown in fig. 3, the specific steps of establishing the ECMrRVESM model based on the error check for the multi-pool correlation vector echo state machine include:

data input and normalization processing:

Training the MrRVESM model to obtain a preliminary prediction result of the electric quantity of the agent electricity purchasing user:

during the training of the MrRVESM model, a time point t is selected after the end of the pool idling _i Obtaining training errors of each time point, and generating a brand-new error sequence; then, each reserve pool is weighted and combined, and the weight is optimized; finally, weather data such as temperature, precipitation and the like which are obviously related with the electricity consumption prediction result are selected, and a unique influence factor sequence is constructed to obtain an electricity preliminary prediction value y _t 。

Constructing a new error training set in which an error sequence and influence factors are fused:

Training and obtaining new error sequence results:

And finishing checking the preliminary power consumption predicted value according to the new error sequence considering the influence factors:

according to the short-term electricity consumption prediction requirement of the agent electricity purchasing user, the performance is improved by adjusting the connection weight and other parameters. Training the model by using marked training data, fusing an error prediction result with an output value of the MrRVESM model by using an error checking algorithm, and constructing a multi-reservoir related vector echo state machine EC-RVMESM based on error result checking. And by checking and correcting the output result, the influence of errors is reduced. And further finishing checking the electric quantity prediction result of the agent electricity purchasing user, and obtaining a final checked prediction result y.

The novel reserve pool construction module further comprises:

a state node definition module in the pool:

a random initialization state node weight module:

Dynamically updating the pool state module:

the process of updating the state node is described as follows:

x(t)＝(1-a)x(t-1)+a×tanh[W _in u(t)+Wx(t-1)+W _back y(t-1)] (1)

input and output modules for setting a reserve pool:

X＝[x(t),x(t+1),...,x(t+N)] ^T (2)

The output after pool mapping is:

y(X)＝W _out X (3)

in which W is _out The weight matrix is output;

repeating training and adjusting the reserve pool module:

The multi-pool correlation vector echo state machine model building module further comprises:

a single pool construction module for updating the state node value in each pool by using a state update equation, processing the state update of the whole pool in one operation by adopting a vectorization mode to generate a plurality of different pools, X ₁ ,X ₂ ,...,X _n Then, weighted combination is carried out to realize updating;

Wherein, the double reservoir is expressed as:

X＝λX ₁ +(1-λ)X ₂ (4)

the multi-reservoir correlation vector echo state machine model outputs are:

y＝ω ^T X+ε (5)

in the above, I ² Representing matrix norm operation and then taking a square value;

Wherein: omega= (omega) ₀ ,ω ₁ ,...,ω _i ,...,ω _N ) ^T The method comprises the steps of carrying out a first treatment on the surface of the Alpha is variance, alpha= (alpha) ₀ ,α ₁ ,...,α _i ,...,α _N ) ^T The method comprises the steps of carrying out a first treatment on the surface of the p (ω|α) represents the conditional probability of ω;is omega _i Is a normal distribution density function of (2);

p(ω|y,α,σ ² )＝N(ω|μ,∑) (9)

Σ＝(σ ^-2 X ^T X+α) ^-1 (10)

μ＝σ ^-2 ΣX ^T y (11)

wherein: gamma ray _i ＝1-α _i Σ _i,i ，Σ _i,i Represents the diagonal elements of the sigma matrix, μ _i Then the value obtained by continually iterating through the update (10);

y _* ＝μ ^T X _* (14)。

the multi-reservoir related vector echo state machine model building module based on error checking further comprises:

for the MrRVESM model training module, in the process of the MrRVESM model training, a time point t is selected after the idling of the reserve pool is finished _i Obtaining training errors of each time point, and generating a brand-new error sequence; then, each reserve pool is weighted and combined, and the weight is optimized; finally, weather data such as temperature, precipitation and the like which are obviously related with the electricity consumption prediction result are selected, and a unique influence factor sequence is constructed to obtain an electricity preliminary prediction value y _t ；

The invention is further illustrated by the following specific examples:

1. data processing and evaluation index

And carrying out simulation analysis by using the actual agent electricity purchasing user electricity consumption data of the power grid in the specific region. The initial data includes historical electricity consumption data and various influence factor data, and covers a period of one year. Historical electricity usage data was sampled at 15 minute intervals for a total of 96 sets of electricity usage data per day. Influencing factors include temperature, humidity, precipitation, and workday type, etc. All raw data is subjected to a normalization process operation prior to model training.

And selecting RVM model, RVESM model, mrRVESM model and ECMrRVESM model for comparison experiment. Firstly, 2,000 neurons are adopted for each pool, the spectrum radius is 0.9, the input of linear combination is converted into nonlinear output, a hyperbolic tangent function is adopted to update state nodes in the embodiment, and the sparsity is set to be 5%. Meanwhile, a particle swarm algorithm is applied to optimize the weights of a plurality of storage pools. And separating the working day from the non-working day, and performing independent prediction to avoid the influence of the switching of the working day and the non-working day on the power consumption prediction result of the agent electricity purchasing user. Next, a part of the 20 days data before the day to be measured is selected as a training set, that is, input data. The remaining data is used for reservation of the pool. When the electricity consumption of the agent electricity purchasing user is predicted, assuming that the current predicted point is at the time t, predicting from historical electricity consumption data and related weather influencing factors at the time t, t+1 and t-1 in the past day (the weight of bad weather is 0, the weight of cloudy days is 0.5, and the weight of sunny days is 1.0). Finally, three evaluation indexes were introduced to evaluate the accuracy of the model, namely, average absolute percentage error EMAPE, maximum error EMAX, and root mean square error RMSE (i.e., root mean square error).

2. Analysis of results

(1) Agent electricity purchasing user electricity quantity prediction

The first step is to use RVM model, RVESM model and MrRVESM model to predict the electricity consumption of agent electricity purchasing user. The validity of the reserve pool structure to replace the kernel function is verified through the prediction of the electricity consumption of a specific workday, and whether the multi-reserve pool structure can further improve the prediction accuracy of the model is explored.

TABLE 3 model predictive error analysis

From the data shown in Table 3-1, it can be observed that the RVESM model has a 0.79% decrease in average absolute error and a 0.58% decrease in maximum error, and the RMSE has a 4.82MW decrease, compared to the RVM model. The results show that excellent results were achieved by using the pool instead of the kernel. Compared with the RVESM model, the average absolute error of the MrRVESM model is reduced by 0.61%, the prediction accuracy is higher, and the multi-reservoir architecture can reflect the dynamic change of data more accurately. The maximum error is reduced by 1.24%, the root mean square error is reduced by 15.94MW, and further stronger stability and smaller prediction bias inside the multi-reservoir are verified.

In order to further verify the effectiveness of the multi-pool correlation vector echo state machine, the power consumption of the same workday is predicted by using two combined kernel correlation vector machine models of RBF+Linear and RBF+Sine. For specific prediction error results, please refer to table 3-2 to obtain detailed data.

TABLE 3-2 analysis of prediction errors for three models

The data shown in table 3-2 can observe that the multi-reservoir correlation vector echo state machine is more excellent in prediction accuracy than the common combined kernel correlation vector machine model. This result further verifies the effectiveness and superiority of the multiple reservoir model. Specifically, the average absolute error of the MrRVESM model was reduced by 0.45% and 0.19% respectively, and the maximum error was reduced by 3.14% and 1.99% respectively, as compared to the other two models. Therefore, it can be observed that the multi-reservoir model has slightly improved prediction accuracy compared with the combined kernel function, and the deviation of the predicted value from the true value is also reduced, so that the prediction accuracy is further improved, and the effectiveness of the multi-reservoir model in the prediction of the electricity consumption of the agent electricity purchasing user is further verified.

(2) Completion of agent electricity purchase prediction result check

And extracting the electricity consumption influence factors and combining the first prediction result error sequence to generate a brand-new training error set. This new set of training errors is trained. And obtaining an error compensation value, thereby obtaining a corresponding new error prediction result. And (3) fusing an error prediction result with an output value of the MrRVESM model by using an error checking algorithm, and constructing a multi-reservoir related vector echo state machine EC-RVMESM based on error result checking to check and correct the output result so as to reduce the influence of errors. And further finishing checking the electric quantity prediction result of the agent electricity purchasing user, and obtaining a final checked prediction result.

The final prediction result analysis of the MrRVESM model without error check on the workday and the EC-RVMESM model based on error result check is shown in tables 3-3:

TABLE 3-3 workday two model prediction result error analysis

According to the data in tables 3-3, the average absolute error of the EC-MrRVESM model relative to the original model was reduced by 0.29%, the maximum error was reduced by 0.50%, and the RMSE was also reduced significantly by 32.53MW. By comparing the EC-RVESM model with the RVESM model, the introduction of the error correction link can be obviously seen to play an important role in improving the accuracy of the model, and the result further verifies the positive influence of the error correction link on the model performance.

(3) Model stability verification

For researching the stability of the prediction model, RVESM, mrRVES and EC-MrRVESM models are selected to predict the electricity consumption of the agent electricity purchasing user for a continuous week, and the prediction result errors of the working days and the rest days are analyzed. Specific data are shown in tables 3-4.

Table 3-4 model prediction error analysis within a week 3

By comparing the one week proxy electricity consumer electricity consumption prediction results in tables 3-4, it can be observed that the EC-gvasm model always performs excellent in terms of accuracy and stability of prediction. Compared with other two models, the average absolute error of the EC-GVESM model is reduced by 1.09 percent and 0.48 percent respectively, and the prediction accuracy is greatly improved; at the same time, the average value of the maximum error is reduced by 3.10% and 0.89%; RMSE is only 36.37MW, stability is significantly improved; good improvement is also obtained for data points with larger prediction errors. To account for the significant differences between the power usage data of the weekday and holiday agent power purchase users, the two types of data are clearly distinguished prior to training. Therefore, when predicting the working day, using the complete working daily electricity quantity data as a training set; and when the rest day is predicted, the complete rest daily electricity consumption data is used as a training set, so that adverse effects on a prediction result caused by the suddenly-reduced rest daily electricity consumption condition are avoided, and the prediction accuracy is remarkably improved.

And selecting the power consumption of one week in spring of a certain place for prediction according to the prediction. However, in the task of predicting the electricity consumption of the agent electricity purchasing user, stability of the predicted result is also an important test factor for different seasons. Compared with spring, the fluctuation of the electricity consumption of the agent electricity purchasing user in summer is larger and is mainly influenced by factors such as an air conditioner, so that the uncertainty of electricity consumption prediction of the agent electricity purchasing user is increased. Therefore, in order to evaluate the prediction performance of the model in different seasons, the power consumption data of the agent electricity purchasing user in the week in summer is selected for prediction analysis. As can be seen from tables 3-5, the EC-MrRVESM model exhibits excellent effects in summer predictions. It can be seen that the EC-MrRVESM model exhibits excellent performance and stability in short-term electricity consumption predictions throughout the seasons, with small overall prediction errors.

Tables 3-5 prediction results of summer weekly electricity consumption

The innovation of the invention is that:

1. according to the invention, the input data is mapped by replacing a kernel function by using the reserve pool structure, and the data characteristics are effectively displayed by establishing a weighted combination mechanism of a plurality of reserve pools, so that a complicated parameter optimization process is avoided;

2. According to the invention, the distribution of the output weights is rapidly calculated by adopting a sparse Bayesian method, so that the time required by solving the output weights of the multi-reserve pool structure is effectively reduced;

3. according to the invention, the error checking structure is used, and in the model training process, the prediction result is corrected by performing secondary training and error checking. And the error prediction result is fused with the output value of the MrRVESM model, so that the power consumption prediction result of the agent power purchase user of the MrRVESM model is checked, and the accuracy of the power consumption prediction result of the agent power purchase user group is effectively improved.

The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

The foregoing description of the preferred embodiments of the invention is not intended to be limiting, but rather is intended to cover all modifications, equivalents, and alternatives falling within the spirit and principles of the invention.

Claims

1. The method for predicting the electric quantity of the agent electricity purchasing user based on the echo state machine and error check is characterized by comprising the following steps of: the method comprises the following steps:

2. The method for predicting the electric quantity of the agent electricity purchasing user based on the echo state machine and the error check according to claim 1, wherein the method comprises the following steps: the specific steps of constructing the novel reserve pool according to the time series data of the historical electricity consumption and the electricity consumption influence factors include:

the process of updating the state node is described as follows:

x(t)＝(1-a)x(t-1)+a×tanh[W _in u(t)+Wx(t-1)+W _back y(t-1)] (1)

in the above expression, it is assumed that x (t) represents the pool state value at the present timeX (t-1) represents the state value of the reserve tank at the previous time, u (t) represents the input value at the current time, and y (t-1)]An output value indicating the previous time, a being the update rate; in this arrangement, W _in Represents the weight matrix of the input layer, W represents the weight matrix connected with the interior of the storage pool, W _back Representing a connection weight matrix of the output layer to the reserve tank; tanh () is an activation function;

X＝[x(t),x(t+1),…,x(t+N)] ^T (2)

The output after pool mapping is:

y(X)＝W _out X (3)

in which W is _out The weight matrix is output;

3. The method for predicting the electric quantity of the agent electricity purchasing user based on the echo state machine and the error check according to claim 1, wherein the method comprises the following steps: the specific steps of constructing the multi-reservoir related vector echo state machine model based on the constructed novel reservoir comprise the following steps:

for the followingEach reservoir uses a state update equation to update state node values inside the reservoir, and the state update of the whole reservoir is processed in one operation in a vectorization mode to generate a plurality of different reservoirs, X ₁ ,X ₂ ,...,X _n Then, weighted combination is carried out to realize updating;

connecting the outputs of the plurality of reservoirs together to form a final output layer, and integrating the outputs of the plurality of reservoirs through a simple connecting operation;

Wherein, the double reservoir is expressed as:

X＝λX ₁ +(1-λ)X ₂ (4)

the multi-reservoir correlation vector echo state machine model outputs are:

y＝ω ^T X+ε (5)

wherein: omega= (omega) ₀ ,ω ₁ ,...,ω _i ,...,ω _N ) ^T The method comprises the steps of carrying out a first treatment on the surface of the Alpha is variance, alpha= (alpha) ₀ ,α ₁ ,...,α _i ,...,α _N ) ^T The method comprises the steps of carrying out a first treatment on the surface of the p (ω|α) represents the conditional probability of ω; n (omega) _i |0,α _i ^-1 ) Is omega _i Is a normal distribution density function of (2);

p(ω|y,α,σ ² )＝N(ω|μ,∑) (9)

∑＝(σ ^-2 X ^T X+α) ^-1 (10)

μ＝σ ^-2 ∑X ^T y (11)

y _* ＝μ ^T X _* (14)。

4. the method for predicting the electric quantity of the agent electricity purchasing user based on the echo state machine and the error check according to claim 1, wherein the method comprises the following steps: the specific steps of establishing the multi-reservoir related vector echo state machine model based on error check, inputting historical electricity consumption data and influence factor data of electricity consumption of different granularities in a matrix form, and finally outputting an electricity prediction result comprise:

Determining the dimension and format of input data, wherein the input data comprises information related to agent electricity purchasing business such as historical electricity purchasing data, market price data, generating capacity data and the like, carrying out normalization processing and daily type separation on the data, and initializing a weight matrix;

during the training of the MrRVESM model, a time point t is selected after the end of the pool idling _i Obtaining training errors of each time point, and generating a brand-new error sequence; then, each reserve pool is weighted and combined, and the weight is optimized; finally, weather data such as temperature, precipitation and the like which are obviously related with the electricity consumption prediction result are selected, and a unique influence factor sequence is constructed to obtain an electricity preliminary prediction value y _t ；

Extracting power consumption influence factors according to the power consumption characteristics of the power consumption of the agent power purchase user; combining the training error sequence with the influence factor sequence to generate a brand-new training error set;

training the new training error set by using a correlation vector machine model RVM, and predicting the error of the current time point after the training is completed to obtain an error compensation value, thereby obtaining a corresponding new error prediction result;

according to the short-term electricity consumption prediction requirement of the agent electricity purchasing user, the performance is improved by adjusting the connection weight and other parameters; training the model by using marked training data, and fusing an error prediction result with an output value of the MrRVESM model by using an error checking algorithm to construct a multi-reservoir related vector echo state machine EC-RVMESM based on error result checking; the output result is checked and corrected, so that the influence of errors is reduced; and further finishing checking the electric quantity prediction result of the agent electricity purchasing user, and obtaining a final checked prediction result y.

5. The utility model provides a proxy electricity purchase user's electric quantity prediction device based on echo state machine and error check which characterized in that: comprising the following steps:

6. The device for predicting the power consumption of a proxy electricity purchasing user based on an echo state machine and error checking according to claim 5, wherein: the novel reserve pool construction module further comprises:

a state node definition module in the pool:

A random initialization state node weight module:

dynamically updating the pool state module:

the process of updating the state node is described as follows:

x(t)＝(1-a)x(t-1)+a×tanh[W _in u(t)+Wx(t-1)+W _back y(t-1)] (1)

input and output modules for setting a reserve pool:

X＝[x(t),x(t+1),...,x(t+N)] ^T (2)

the output after pool mapping is:

y(X)＝W _out X (3)

in which W is _out The weight matrix is output;

repeating training and adjusting the reserve pool module:

7. The device for predicting the power consumption of a proxy electricity purchasing user based on an echo state machine and error checking according to claim 5, wherein: the multi-pool correlation vector echo state machine model building module further comprises:

a single pool construction module for updating the state node value in each pool by using a state update equation, processing the state update of the whole pool in one operation by adopting a vectorization mode to generate a plurality of different pools, X ₁ ,X ₂ ,...,X _n Then, weighting group is carried outCombining to realize updating;

wherein, the double reservoir is expressed as:

X＝λX ₁ +(1-λ)X ₂ (4)

The multi-reservoir correlation vector echo state machine model outputs are:

y＝ω ^T X+ε (5)

in the following.|| ² Representing matrix norm operation and then taking a square value;

p(ω|y,α,σ ² )＝N(ω|μ,∑) (9)

∑＝(σ ^-2 X ^T X+α) ^-1 (10)

μ＝σ ^-2 ∑X ^T y (11)

y _* ＝μ ^T X _* (14)。

8. the device for predicting proxy electricity purchasing user electric quantity based on an echo state machine and error checking according to claim 5, wherein the multi-reservoir related vector echo state machine model building module based on error checking further comprises:

9. A computer device comprising a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the power prediction method of any of claims 1-4.

10. A computer readable storage medium storing a computer program which, when executed by a processor, causes the processor to perform the steps of the power prediction method of any one of claims 1-4.