CN106709820A - Power system load prediction method and device based on deep belief network - Google Patents

Abstract

The embodiment of the invention provides a method and a device for predicting a load of a power system based on a deep belief network, relates to the field of power systems, and can improve convergence speed and reduce prediction errors. The specific scheme comprises the following steps: acquiring a training sample and a test sample; constructing an energy function of the RBM model; training the at least one hidden layer and the visible layer by utilizing the training sample to obtain the weight of the training sample between the nodes of the at least one hidden layer and the visible layer; and inputting the output data obtained by the training sample and the test sample into the trained DBN to obtain a predicted value of the power system load. The method is used for predicting the load of the power system.

Description

A method and device for power system load forecasting based on deep belief network

technical field

Embodiments of the present invention relate to the field of power systems, and in particular to a method and device for power system load forecasting based on a deep belief network.

Background technique

Power system load forecasting is an important part of power system planning and the basis of power system economic operation. It starts from the known power demand and fully considers the influence of political, economic, climate and other related factors to predict future power demand. . Accurate load forecasting data is helpful for power grid dispatching control and safe operation, formulating reasonable power supply construction planning and improving the economic and social benefits of the power system. In load forecasting, methods such as regression model, time series forecasting technology and gray theory forecasting technology are often used. In recent years, with the rise of artificial neural network research, the use of neural network for power system load forecasting has greatly reduced the forecasting error, which has aroused great attention of load forecasting workers.

In the prior art, the traditional neural network is used for load forecasting of the power system. However, the traditional neural network is a typical global approximation network, and one or more weights of the network have an influence on each output; on the other hand, the network has randomness in determining the weights, resulting in the input after each training , The relationship between the output is uncertain, and the prediction results are different. It can be seen that the solutions in the prior art have the problems of slow convergence speed and large prediction error.

Contents of the invention

Embodiments of the present invention provide a power system load forecasting method and device based on a deep belief network, which can improve convergence speed and reduce forecasting errors.

In order to achieve the above purpose, the embodiments of the present application adopt the following technical solutions:

In the first aspect, a DBN-based power system load forecasting method is provided. The DBN is composed of a multi-layer restricted Boltzmann machine RBM, including at least one hidden layer and one visible layer. The power system load forecasting method includes:

Obtain training samples and test samples; construct the energy function of the RBM model;

Using the training samples to train the at least one hidden layer and visible layer layer by layer, to obtain the weight of the training samples between the at least one hidden layer and visible layer nodes;

The output data obtained from the training samples and the test samples are input into the trained DBN to obtain the predicted value of the power system load.

In the second aspect, a DBN-based power system load forecasting device is provided, which is used to implement the method provided in the first aspect.

The power system load forecasting method and device based on the deep belief network provided by the embodiment of the present invention uses a multi-layer restricted Boltzmann machine to form a deep belief network, including several hidden layers and a visible layer, which can be achieved by using hierarchical The supervised greedy pre-training method pre-trains RBM hierarchically, and uses the obtained results as the initial value of the supervised learning training probability model, thereby greatly improving the learning performance, increasing the convergence speed, and reducing the prediction error.

Description of drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the accompanying drawings that need to be used in the description of the embodiments. Obviously, the accompanying drawings in the following description are only of the present invention. For some embodiments, those of ordinary skill in the art can also obtain other drawings based on these drawings without any creative effort.

Fig. 1 is a schematic flow chart of a power system load forecasting method based on a deep belief network provided by an embodiment of the present invention;

Fig. 2 is a schematic diagram illustrating when RBM includes a visible layer and a hidden layer;

Fig. 3 is a schematic diagram illustrating the optimal error recognition rate under different structures of three models in an embodiment of the present invention;

In the embodiment of the present invention in Fig. 4, the curve graph of the average absolute percentage error changing with the number of iterations;

FIG. 5 is a schematic structural diagram of a DBN-based power system load forecasting device provided by an embodiment of the present invention.

detailed description

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

Aiming at the slow convergence speed and large prediction error of traditional neural network load forecasting, this paper proposes a power system load forecasting scheme based on Deep Belief Network (English full name: Deep Belief Network, English abbreviation: DBN), which is shown in Figure 1 , the method includes the following steps:

101. Obtain training samples and test samples.

102. Construct the energy function of the RBM model.

103. Use the training samples to train at least one hidden layer and visible layer layer by layer, and obtain weights of the training samples between nodes in the at least one hidden layer and the visible layer.

104. Input the output data obtained from the training samples and the test samples into the trained DBN to obtain a predicted value of the power system load.

In a specific embodiment of the present invention, the DBN is composed of a multi-layer Restricted Boltzmann Machine (English full name: RestrictedBoltzmann Machine, RBM), including at least one hidden layer and one visible layer. Layer pre-training RBM, and the obtained result is used as the initial value of the supervised learning training probability model, thereby greatly improving the learning performance, as follows:

1. Data preprocessing

S1. Preprocess the training samples:

Assume that all training samples are X={X ₁ ,X ₂ ,...,X _{M }} , where Xi represents a set of load data, and M represents the number of load data. For input samples, the algorithm adopts 4 steps for regression One chemical:

① Calculate the mean:

Among them, u represents the mean of the sample.

② Calculate the variance:

Among them, δ represents the variance of the sample.

③Albinism:

Among them, X _i ' represents the whitening parameter of the sample data.

④Normalization:

Among them, X _{i, n} represent the normalization parameters of the sample data;

S2. Preprocessing the test samples:

The preprocessing of the test sample needs to be whitened according to the mean and variance of the training sample, and then uniformly normalized to between 0 and 1 according to the maximum and minimum values of the training sample. Assuming that a single test sample is T, the normalization step is :

① Whitening:

Among them, T' represents the whitening parameter of the test sample data, T is a single test sample, u represents the mean of the training samples, and δ represents the variance of the training samples.

②Normalization:

Among them, T _n represents the normalization parameter of the test sample data.

Here, the reason for whitening is because there is a large correlation between adjacent elements of natural data, so the redundancy of data can be reduced through whitening, similar to principal component analysis (English full name: Principal Component Analysis, English name: PCA) dimensionality reduction.

2. The basic principle and composition of deep belief network

DBN is an energy model with a very powerful feature learning ability, and it is a deep network that has been studied very early. The learning method was proposed by Hinton et al. The essence is to learn more useful features by building a machine learning model with many hidden layers and a large amount of training data, so as to finally improve the accuracy of classification or prediction.

DBN can be regarded as a complex neural network composed of multi-layer RBM, which includes a visible layer and several hidden layers, and uses the method of contrast divergence to train the pre-weight value of RBM. The deep neural network uses the hierarchical unsupervised greedy pre-training method to pre-train the RBM hierarchically, and the obtained results are used as the initial value of the supervised learning training probability model, and the learning performance is greatly improved.

RBM is a generative random neural network, which is used to learn the probability distribution of input data. It can be trained in both supervised and unsupervised ways as needed. It is widely used in many aspects. As shown in Figure 2, the RBM has two layers of networks, which are respectively expressed as: vector h=(h ₁ ,h ₂ ,…h _m ) represents the hidden layer unit, and there are m elements in total; vector v=(v ₁ ,v ₂ ,…, v _n ) represents the visible layer, with a total of n elements. Because there is no connection between the internal elements of the layer, each element in the layer is independent of each other, that is:

p(h|v)＝p(h ₁ |v)p(h ₂ |v)...p(h _m |v);

p(v|h)=p(v ₁ |h) p(v ₂ |h)...p(v _n |h).

Generally, the distribution of visible layer and hidden layer unit satisfies the Bernoulli form, then:

It can be seen that the probability distribution of the entire model can be obtained through the probability distribution of each layer, so that unsupervised training can be performed layer by layer. The output of the lower layer RBM is used as the input of the upper layer RBM. Up to the last hidden layer, the last hidden layer is input to the softmax classifier, and then the estimated error is calculated for reverse fine-tuning.

S3, constructing the energy function of the RBM model:

The energy function of the RBM model satisfying the Bernoulli distribution of the hidden layer and the visible layer is defined as:

Among them, a _i and b _j represent bias items, and w _ij represents the connection weight between visible units and hidden layer units. θ=(a,b,w) is the weight parameter of the RBM model, the vector h=(h ₁ ,h ₂ ,…h _m ) represents the hidden layer unit, there are m elements in total, the vector v=(v ₁ ,v ₂ ,…,v _n ) represent the visible layer, with a total of n elements.

S4. Calculate the joint probability distribution of visible layer and hidden layer units:

Given the model parameters, the joint probability distribution of visible and hidden layer units can be expressed as:

in, is a normalization factor, -E(v,h) is the energy function of the RBM model of the hidden layer and the visible layer, the vector h=(h ₁ ,h ₂ ,…h _m ) represents the hidden layer unit, and there are m elements in total .

According to formula (7) and formula (8), it can be deduced that:

In the formula,

S5. Calculate the edge distribution of the visible layer and the hidden layer:

According to the joint probability distribution P(v,h) of the visible layer and the hidden layer, the marginal distribution can be obtained:

in, is a normalization factor, E(v,h) is the energy function of the RBM model of the hidden layer and the visible layer, and the vector h=(h ₁ ,h ₂ ,…h _m ) represents the hidden layer unit, with m elements in total.

S6, constructing the logarithmic likelihood function is:

Among them, l is the number of training samples, θ=(a,b,w) is the weight parameter of the RBM model, P(v _i ) is the marginal distribution of the visible layer and the hidden layer.

S7. Calculating the gradient of the logarithmic likelihood function:

According to the gradient descent method to derive the logarithmic likelihood function, we can get:

Organized to get:

Among them, lnL(θ) is the logarithmic likelihood function, l is the number of training samples, θ=(a,b,w) is the weight parameter of the RBM model, P(h|v _i ) is the visible layer and the hidden layer The marginal distribution of , E(v,h) is the energy function of the RBM model of the hidden layer and the visible layer, P(h,v _i ) is the joint probability distribution of the visible layer and hidden layer units.

3. The process of predicting power system load using deep belief network

For deep learning, the idea is to stack multiple layers, that is to say, the output of this layer is used as the input of the next layer. For example, there is a system S, which has n layers (S1,...Sn), its input is I, and its output is O, which can be expressed as: I=>S1=>S2=>.....=>Sn=>O, By adjusting the parameters in the system so that the output O is still the input I, then a series of hierarchical features of the input I can be automatically obtained, namely S1,...,Sn. In this way, hierarchical expression of input information can be realized. The process of using deep belief network to predict power system load is essentially a deep learning process.

The process of using deep belief network to predict power system load is as follows: using the method of contrastive divergence to train the pre-weight value of RBM. Through analysis, the algorithm uses three hidden layers, and adopts hierarchical unsupervised greedy pre-training method to pre-train RBM hierarchically. The obtained results are used as the initial value of the supervised learning training probability model, thereby improving the learning performance. The simulation data show that, compared with the traditional neural network algorithm, the algorithm converges faster and the prediction error is smaller.

① Use the method of contrast divergence to train the pre-weight value of RBM. KL-divergence is an asymmetric measure, so the values of KL(Q||P) and KL(P||Q) are different. The greater the difference between the two distributions, the greater the value of the KL-divergence. The maximum value of the log-likelihood function of the RBM model eventually evolves to calculate the difference between the KL-divergence of two probability distributions:

KL(Q||P)-KL(P _k ||P) (Equation 15)

Among them, Q is the prior distribution, and P _k is the distribution after k-step Gibbs sampling. If the Gibbs chain reaches a steady state (since the initial state v ⁽⁰⁾ = v, it is already a steady state), then P _k =P, that is, KL(P _k ||P)=0. In this way, the estimation error obtained by the contrastive divergence algorithm is equal to 0.

②The obtained result is used as the initial value of the supervised learning training probability model, and the relative entropy, also known as KL-divergence, is often used to measure the distance between two probability distributions. The KL-divergence of two probability Q and P distributions in state space is defined as follows:

KL(Q||P) is the KL-divergence of two probability distributions Q and P in state space.

③Use the obtained pre-weight value and initial value as input data, pass it to the next layer, and then train in the next layer. Train one layer of network at a time, layer by layer. Specifically, the first layer is first trained with uncalibrated data, and the parameters of the first layer are first learned during training (this layer can be regarded as a hidden layer of a three-layer neural network that minimizes the difference between the output and the input). After the n-1th layer is obtained, the output of the n-1 layer is used as the input of the nth layer, and the nth layer is trained to obtain the parameters of each layer respectively. Therefore, the training process is an iterative process. When all layers are trained, use the wake-sleep algorithm for tuning.

Wake stage: the cognitive process, which generates an abstract representation (node state) of each layer through external features and upward weights (cognitive weights), and uses gradient descent to modify the downlink weights between layers (generated weights).

Sleep stage: the generation process, through the top-level representation and downward weights, the underlying state is generated, and the upward weights between layers are modified at the same time.

④ After pre-training, DBN can use the BP algorithm to adjust the discriminative performance by using the labeled data. Here, a label set will be appended to the top layer (generalizing associative memory) to obtain a classification facet of the network through a bottom-up, learned recognition weight. This performance will be better than the network trained by pure BP algorithm. This can be explained intuitively. The BP algorithm of DBNs only needs to perform a local search on the weight parameter space. Compared with the forward neural network, the training is faster and the convergence time is also less.

⑤When the training and tuning steps are over, at the top two layers (ie, the last hidden layer and the visible layer), the weights are connected together so that the output of the lower layer will provide a reference clue or link to the top layer, This way the top layer will associate it with the contents of its memory. Thus, the load value prediction of the electric power system is performed.

In a specific implementation, S8: Calculate the KL-divergence of the two probabilities Q and P of the visible layer and the hidden layer in the state space according to Formula 16; S9: Calculate the logarithmic likelihood of the RBM model according to Formula 15 The maximum value of the function eventually evolves to calculate the difference of the KL-divergence of two probability distributions Q and P. S10: S1-S9 is to carry out hierarchical training on the training samples through the RBM model, and obtain the weight value of the training sample between the visible layer node and the hidden layer node. The weight value is input from the hidden layer to the visible layer, and the visible layer will remember the weight value The content and the load prediction value of the training samples are obtained, which are used in the load prediction process of the power system. S11, the test sample plus the average load value, maximum load value, average temperature, and the predicted daily average temperature of the numerical weather forecast of the training sample obtained in S10 are used as the input data of the deep neural network algorithm, through the calculation of S1-S9, the output data That is to predict the load value of the power system.

4. Experimental settings

The load test used the load data of East Slovakia Electric Power Company from January to October 2007 provided by EUNITE competition as training data, and the data from November to December was used as forecast comparison data. The input data of the neural network includes: 48 sampling points of the previous day (average load data every half hour), the average temperature of the previous day, the peak load of the previous day, the valley value of the load of the previous day, the average value of the previous day, and the predicted daily average temperature . The prediction algorithm predicts 48 points of data for the next day at one time.

In the application of load forecasting, load forecasting data is generally required every half hour within 0 hours to 24 hours. The algorithm takes a total of 48 historical data from 0:00 to 24:00 of the day before the forecast day, plus the average load value and maximum load of the day before the forecast day. The daily average temperature predicted by the numerical weather forecast is used as the input data of the deep neural network algorithm. The output data is the load data every half hour from 0:00 to 24:00 on the forecast day.

In order to compare the performance of the algorithm, BP neural network, self-organizing fuzzy neural network (English full name: Self-organizing Fuzzy Neural Network, English abbreviation: SOFNN) and deep belief network are selected for comparison. BP neural network has good characteristics of capturing nonlinear laws. When there are enough neurons in the hidden layer, the 3-layer perceptron model can realize the approximation of any nonlinear function. The main advantage of the SOFNN algorithm is that it can automatically determine the network structure and give model parameters, which has good prediction accuracy.

The evaluation standard is the mean absolute percentage error (MAPE), which is defined as follows:

Among them, D _i and are the real value and predicted value of the maximum load on the i-th day of a certain month in 1997, respectively, and n is the number of days in a certain month in 1997.

5. Performance Evaluation

(1) Influence of network structure on algorithm performance

First of all, the experiment studies the influence of the network structure (hidden layer, that is, the number of neurons) on the prediction effect, taking the load data from January to October 2007 as the training data, and the load data in November as the test data. The network structure uses 1, 2, and 3 hidden layers respectively. The number of neurons was 20, 50, 100, 200, 400, respectively. Here, for the convenience of comparison, the number of neurons in each layer is the same. For the initial parameter setting, the selection of the following parameters will be manually selected from these ranges to obtain the optimal recognition rate: BP learning rate (0.1, 0.05, 0.02, 0.01, 0.005), pre-training learning rate (0.01, 0.005, 0.002, 0.001). Figure 3 shows the optimal false recognition rates of the three models under different structures. It can be seen from Figure 3 that when the number of hidden layers is small or the number of neurons is small, the performance of the BP network without pre-training is better. When there is only one hidden layer and the number of neurons reaches 200, the error of SOFNN is The rate is comparable to that of BP, and when the number of hidden layers is 2 and 3, and the number of neurons reaches 100 and 50, the performance of SOFNN will be close to and exceed that of BP. The same is true for DBN. It can be found that the performance of SOFNN is better than that of DBN when the number of neurons is small, and the opposite is true when the number of neurons is large. And when the number of hidden layers is the same, because of the overfitting problem. The performance of the BP network does not increase with the number of nodes, but decreases when the number of nodes is too large, while the depth model has been increasing and tends to be stable, which shows that the performance of the pre-trained deep network increases with the expansion of the network size (including The number of hidden layers increases or the number of nodes increases) is gradually optimized. It can be seen that too few hidden layers and hidden nodes will reduce the performance of the deep model. The reason can be explained as follows. The function of the pre-training model is to extract the core features of the input features. Due to the limitation of the sparsity condition, assuming that the number of neurons is too small, only a small number of neurons are activated for the input of some input samples, while These features cannot represent the original input, so some information is lost, resulting in a decrease in performance. Although the larger the network size, the better the performance of the deep model, but at the same time the training time is longer, so a trade-off between performance and training time is required.

(2) Algorithm convergence comparison

Figure 4 is the curve of the mean absolute percentage error versus the number of iterations. The algorithm adopts a model with three hidden layers, 100 neurons in each layer, and 100 neurons in each layer. The learning rate is set to 0.01 during pre-training. When fine-tuning, the learning rate is 0.1, and the coefficient of the momentum item is 0.5. The load data from January to November 2007 is used as training data, and the load data from December is used as test data. Compare the average absolute percentage error changes of BP, SOFNN, and DBN under 1000 iterations.

Analyzing Figure 4, it can be seen that as the number of iterations increases, the error rates of the three models gradually decrease. This is because the distribution of network parameters is getting closer to the minimum point as the number of training increases. However, after reaching a certain number of training times, the error rate of the BP network fluctuates and tends to increase gradually, while the error rate of the pre-trained DBN network decreases steadily, which indirectly proves that pre-training can make the distribution area of the initial parameters of the network closer to minimum point, and effectively avoid local shocks.

Load forecasting provides an important basis for distribution network management decision-making and operation mode. The invention proposes a load forecasting algorithm using a deep belief network to improve the problems of slow learning rate and low forecasting efficiency of the existing neural network algorithm. The simulation results show that compared with the traditional neural network algorithm, the power load forecasting effect of the algorithm based on the deep belief network is obvious.

Embodiments of the present invention also provide a DBN-based power system load forecasting device, which is used to implement the power system load forecasting method described in the above embodiments. As shown in FIG. 5 , the power system load forecasting device includes:

Data processing unit 501, for obtaining training sample and test sample; Construct the energy function of RBM model;

A training unit 502, configured to use training samples to train at least one hidden layer and visible layer layer by layer, to obtain weights of training samples between at least one hidden layer and visible layer nodes;

The prediction unit 503 is configured to input the output data obtained from the training samples and the test samples into the trained DBN to obtain a predicted value of the power system load.

Optionally, the data processing unit 501 is further configured to normalize the training samples X={X ₁ , X ₂ ,...,X _{M }} composed of load data Xi; wherein Xi _represents a set of load Data, M represents the number of load data; it is also used to whiten the test samples according to the mean and variance of the training samples, and uniformly normalize them between 0 and 1 according to the maximum and minimum values of the training samples.

Optionally, the distribution of the visible and hidden layer units satisfies the Bernoulli form;

The energy function of the RBM model satisfying the Bernoulli distribution of the visible layer and the hidden layer defined by the data processing unit 501 is:

Among them, a _i and b _j represent bias items, and w _ij represents the connection weight between visible units and hidden layer units. θ=(a,b,w) is the weight parameter of the RBM model, the vector h=(h ₁ ,h ₂ ,…h _m ) represents the hidden layer unit, there are m elements in total, the vector v=(v ₁ ,v ₂ ,…,v _n ) represent the visible layer, with a total of n elements.

Optionally, the training unit 502 is specifically used to train the pre-weight value of the RBM using a contrastive divergence device, and use the obtained result as the initial value of the supervised learning training probability model; use the obtained pre-weight value and initial value as input data , passed to the next layer of network, and then trained in the next layer of network; each time a layer of network is trained, at least one hidden layer and visible layer are trained layer by layer.

Optionally, the output data obtained from the training samples at least includes the average load value, the maximum load value, the average temperature of the training samples, and the predicted daily average temperature of the numerical weather prediction.

The power system load forecasting method and device based on the deep belief network provided by the embodiment of the present invention uses a multi-layer restricted Boltzmann machine to form a deep belief network, including several hidden layers and a visible layer, which can be achieved by using hierarchical The supervised greedy pre-training method layeredly pre-trains RBM, and uses the obtained result as the initial value of the supervised learning training probability model, thereby greatly improving the learning performance, increasing the convergence speed, reducing the prediction error, and improving the learning rate of the existing neural network algorithm. Slow and low prediction efficiency. The simulation results show that compared with the traditional neural network algorithm, the power load forecasting effect of the algorithm based on the deep belief network is obvious.

The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person familiar with the technical field can easily think of changes or replacements within the technical scope disclosed in the present invention, and should cover Within the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims (10)

1. A power system load prediction method based on a Deep Belief Network (DBN), wherein the DBN is composed of a plurality of layers of Restricted Boltzmann Machines (RBMs) and comprises at least one hidden layer and one visible layer, and the power system load prediction method comprises the following steps:

acquiring a training sample and a test sample; constructing an energy function of the RBM model;

training the at least one hidden layer and the visible layer by utilizing the training sample to obtain the weight of the training sample between the nodes of the at least one hidden layer and the visible layer;

and inputting the output data obtained by the training sample and the test sample into the trained DBN to obtain a predicted value of the power system load.

2. The method of claim 1, wherein the constructing the RBM model energy function is preceded by:

for the load data X_iFormed training sample X ═ X₁,X₂,...,X_MNormalizing; wherein X_iRepresenting a group of load data, and M representing the number of the load data;

and whitening the test sample according to the mean value and the variance of the training sample, and uniformly normalizing the test sample to be between 0 and 1 according to the maximum value and the minimum value of the training sample.

3. The power system load prediction method of claim 1, wherein the distribution of visible layer and hidden layer units satisfies bernoulli form; the energy function for constructing the RBM model comprises the following steps:

the energy function of the RBM model satisfying bernoulli distribution for the visible and hidden layers is defined as:

$E(v,h)=-\underset{i=1}{\overset{n}{\Σ}}\underset{j=1}{\overset{m}{\Σ}}{v}_i{w}_{ij}{h}_j-\underset{i=1}{\overset{n}{\Σ}}{a}_i{v}_i-\underset{j=1}{\overset{m}{\Σ}}{b}_j{h}_j;$

wherein, a_iAnd b_jRepresents a bias term, w_ijRepresenting the connection weight between the visible unit and the hidden layer unit, theta ═ a, b, w are weight parameters of the RBM model, and the vector h ═ h₁,h₂,…h_m) Represents a hidden layer unit, has m elements in total, and has a vector v ═ v₁,v₂,…,v_n) Representing a visible layer, for a total of n elements.

4. The method according to claim 3, wherein the training the at least one hidden layer and the visible layer by using the training samples layer by layer to obtain the weight of the training samples between the at least one hidden layer and the visible layer node comprises:

training the preweight of the RBM by adopting a divergence contrast method, and taking the obtained result as an initial value of a probability model for supervised learning training;

the obtained preset value and the initial value are used as input data and transmitted to a next layer of network, and training is carried out in the next layer of network;

and training the network one layer at a time, and training the at least one hidden layer and the visible layer by layer.

5. The power system load prediction method according to claim 3,

the output data obtained from the training samples at least comprise the average load value, the maximum load value, the average air temperature and the predicted daily average air temperature of the numerical weather forecast of the training samples.

6. An electric power system load prediction device based on a Deep Belief Network (DBN), wherein the DBN is composed of a plurality of layers of Restricted Boltzmann Machines (RBMs) and comprises at least one hidden layer and one visible layer, and the electric power system load prediction device comprises:

the data processing unit is used for acquiring a training sample and a test sample; constructing an energy function of the RBM model;

the training unit is used for training the at least one hidden layer and the visible layer by utilizing the training samples to obtain the weight of the training samples between the nodes of the at least one hidden layer and the visible layer;

and the prediction unit is used for inputting the output data obtained by the training sample and the test sample into the trained DBN to obtain a predicted value of the power system load.

7. The power system load prediction device of claim 6,

the data processing unit is also used for processing the load data X_iFormed training sample X ═ X₁,X₂,...,X_MNormalizing; wherein X_iRepresenting a group of load data, and M representing the number of the load data; and the device is also used for whitening the test sample according to the mean value and the variance of the training sample and uniformly normalizing the test sample to be between 0 and 1 according to the maximum value and the minimum value of the training sample.

8. The power system load prediction device of claim 6,

the distribution of the visible layer and the hidden layer units meets the Bernoulli form;

the energy function of the RBM model satisfying Bernoulli distribution of the visible layer and the hidden layer defined by the data processing unit is as follows:

$E(v,h)=-\underset{i=1}{\overset{n}{\Σ}}\underset{j=1}{\overset{m}{\Σ}}{v}_i{w}_{ij}{h}_j-\underset{i=1}{\overset{n}{\Σ}}{a}_i{v}_i-\underset{j=1}{\overset{m}{\Σ}}{b}_j{h}_j;$

wherein, a_iAnd b_jRepresents a bias term, w_ijRepresenting the connection weight between the visible unit and the hidden layer unit, theta ═ a, b, w are weight parameters of the RBM model, and the vector h ═ h₁,h₂,…h_m) Represents a hidden layer unit, has m elements in total, and has a vector v ═ v₁,v₂,…,v_n) Representing a visible layer, for a total of n elements.

9. The power system load prediction device according to claim 8,

the training unit is specifically used for training the preset value of the RBM by adopting a device for comparing divergence, and taking the obtained result as an initial value of the probability model for supervised learning training; the obtained preset value and the initial value are used as input data and transmitted to a next layer of network, and training is carried out in the next layer of network; and training the network one layer at a time, and training the at least one hidden layer and the visible layer by layer.

10. The power system load prediction device according to claim 8,

the output data obtained from the training samples at least comprise the average load value, the maximum load value, the average air temperature and the predicted daily average air temperature of the numerical weather forecast of the training samples.