Short-term load probability density prediction method, device and system based on quantile regression

Abstract

The present invention discloses a short-term load probability density prediction method, device and system based on quantile regression, including obtaining a DL-LSTM-A deep network model, wherein the DL-LSTM-A deep network model includes a plurality of double-layer LSTM network cells and an attention mechanism module, wherein the output ends of each double-layer LSTM network cell are respectively connected to the attention mechanism module; the attention mechanism module performs weighted summation on the outputs of each double-layer LSTM network cell; the double-layer LSTM network cell includes two LSTM cells connected in sequence; the parameters to be optimized in the DL-LSTM-A network are obtained by training with training data, and the optimized DL-LSTM-A deep network model is obtained; the influencing factors of the acquired power load consumption are input into the optimized DL-LSTM-A deep network model to obtain quantiles; the quantiles are processed by a non-parametric density estimation method to obtain a final load probability density function. The present invention can provide a more accurate long-term power load probability density prediction.

Description

Technical Field

[0001] The present invention belongs to the field of smart grid power load prediction, and specifically relates to a short-term load probability density prediction method, device and system based on quantile regression. Background Art

[0002] Compared with the traditional power load point prediction process, the probability density prediction of load can give features such as quantiles, prediction intervals or probability density functions to describe the uncertainty in power load, which has received continuous attention from academia and industry. In addition, from the perspective of time span, the prediction range can range from minutes, hours to days, months and years. Different demands in the power grid correspond to different prediction time spans. Since short-term power load forecasting can provide accurate power load forecasting, it has received widespread attention.

[0003] The short-term probability density prediction of power load is generally divided into two steps. The first step is to use quantile regression analysis to obtain the quantile of power load, and the second step is to use kernel density estimation to obtain the probability density function. Linear quantile regression uses linear functions to describe the relationship between power load and parameters, which is a relatively basic method. Since power load usually presents nonlinear characteristics, the results predicted by linear functions are not very good. At this stage, nonlinear prediction methods, such as those based on SVM models and CNN-GRU models, cannot effectively describe the high-order characteristics of complex time series data, and the prediction accuracy needs to be improved. Therefore, improving the accuracy of power load prediction is an urgent problem to be solved. Summary of the invention

[0004] In view of the above problems, the present invention proposes a short-term load probability density prediction method, device and system based on quantile regression, which can provide a more accurate long-term power load probability density prediction.

[0005] In order to achieve the above technical objectives and the above technical effects, the present invention is implemented through the following technical solutions:

[0006] In a first aspect, the present invention provides a short-term load probability density forecasting method based on quantile regression, comprising:

[0007] Obtain a DL-LSTM-A deep network model, wherein the DL-LSTM-A deep network model includes a plurality of double-layer LSTM network cells and an attention mechanism module, wherein the output ends of each double-layer LSTM network cell are respectively connected to the attention mechanism module; the attention mechanism module performs weighted summation on the outputs of each double-layer LSTM network cell; and the double-layer LSTM network cell includes two LSTM cells connected in sequence;

[0008] Use the training data to train and obtain the various parameters to be optimized in the DL-LSTM-A network, and obtain the optimized DL-LSTM-A deep network model;

[0009] The acquired influencing factors of power load consumption are input into the optimized DL-LSTM-A deep network model to obtain quantiles;

[0010] The non-parametric density estimation method is used to process the quantiles to obtain the final load probability density function and complete the short-term load probability density forecast based on quantile regression.

[0011] Optionally, the method for obtaining the quantile includes:

[0012] The calculation formula for the power load forecast value Y is defined as:

[0013] Y=f(h, d, ty, s, te, n, l, Y o )

[0014] Where h∈{1, 2, ..., 24} is the number of hours in a day, d∈{1, 2, ..., 365} refers to a day in a year; ty is the type of day, ty=1 refers to a working day, ty=0 refers to a rest day; s refers to the season, {1, 2, 3, 4} represents the four seasons of spring, summer, autumn and winter respectively; te refers to the temperature at the current time, described in degrees Celsius; n refers to the number of users served by the current terminal, Y o is the historical load value, f(·) is a nonlinear complex function used to describe the relationship between power load and various influencing factors;

[0015] The distribution function of the power load forecast value Y is defined as:

[0016] F(y)=P(Y≤y)

[0017] Then the minimum y value that satisfies F(y)≥τ is the τth quantile of variable y, and the representation of the τth quantile is:

[0018] Q(τ)=inf(y|F(y)≥τ).

[0019] Optionally, the output expression of the attention mechanism module is:

[0020]

[0021] Among them, n+1 refers to the length of the cell sequence in the proposed attention mechanism, β k refers to the attention value at the t-(k-1)th cell, h t-(k-1) Refers to the output value of the t-(k-1)th cell, H t Refers to the output of the attention mechanism, normalized using the norm, and the attention value βk It can be calculated using the following formula:

[0022]

[0023] Among them, ||·||2 refers to the second norm of the quantity, s k Used to describe the importance of cells, and then H t The output value obtained by inputting the fully connected layer is the output of the DL-LSTM-A network model, which is the predicted value of the power load.

[0024] Optionally, assume that after passing through the DL-LSTM-A network, the power load forecast value Y is related to the network parameters W, V, b and input X = [h, d, ty, s, te, n, l, Y o ] becomes Y=f DL-LSTM-A (X, W, V, b), then the parameters to be optimized in the DL-LSTM-A network obtained by training with the training data include W, V, b, specifically:

[0025] The design optimization objective function is:

[0026]

[0027] where ||·|| F refers to the matrix Frobenius norm, ||·||2 refers to the 2-norm, λ1, λ2, λ3 are three regularization variables, and the function f loss for:

[0028]

[0029] Among them, τ is the quantile, N is the number of samples, and Y i Refers to the output power sample, X i and X′ i All refer to samples under the quantile τ condition, i|Y i ≥f DL-LSTM-A (X i ′, W, V, b) indicates that the i-th response variable is greater than or equal to the regression estimate, i|Y i <f DL-LSTM-A (X i ′, W, V, b) indicates that the i-th response variable is less than the regression estimate, the quantile τ∈(0,1), and the function ρ τ (u)=u(τ-I(u)), I(u) is called the indicator function, u is the independent variable,

[0030] Based on the optimization objective function, the parameters are optimized in combination with the adam optimization algorithm to obtain the parameters to be optimized in the DL-LSTM-A network.

[0031] Optionally, the probability density function is:

[0032]

[0033] in, is the Gaussian kernel function, x and x i is the independent variable, σ is the smoothing parameter, and the parameters in the Gaussian function are obtained using the cross-validation criterion.

[0034] Optionally, each LSTM cell includes four gates, the first of which is called the forget gate. The output of the first gate is f t and input h t-1 , X t And the first door parameter W f , b f The relationship between them is:

[0035] f t =sigmoid(W f [h t-1 , X t ]+b f )

[0036] sigmoid(·) is the sigmoid function; the second and third layers are called the input layer and tanh layer, respectively. The Tanh layer introduces a new candidate value. t and the second layer input h t-1 , X t , the second layer parameter W i 、b i The relationship between them, and the third layer output With the third layer input h t-1 , X t and the third layer network parameter W C 、b C The relationship between them is described by the following formula:

[0037] i t =sigmoid(W i [h t-1 , X t ]+b i )

[0038]

[0039] tanh(·) is the tanh function, with f t 、i t and the output C of the previous LSTM cell t-1 After that, the two outputs C of the LSTM cell t and h tIt is obtained from the following formula:

[0040]

[0041] O t =sigmoid(W O [h t-1 , X t ]+b O ) t

[0042] h t =O t tanh(C t ).

[0043] Where W O 、b O is the parameter of the fourth layer, O t Intermediate variable, h t-1 , X t is the input parameter.

[0044] In a second aspect, the present invention provides a short-term load probability density prediction device based on quantile regression, comprising:

[0045] An acquisition module is used to acquire a DL-LSTM-A deep network model, wherein the DL-LSTM-A deep network model includes a plurality of double-layer LSTM network cells and an attention mechanism module, wherein the output ends of each double-layer LSTM network cell are respectively connected to the attention mechanism module; the attention mechanism module performs weighted summation on the outputs of each double-layer LSTM network cell; and the double-layer LSTM network cell includes two LSTM cells connected in sequence;

[0046] The optimization module is used to obtain various parameters to be optimized in the DL-LSTM-A network by using the training data to obtain the optimized DL-LSTM-A deep network model;

[0047] An execution module is used to input the acquired influencing factors of power load consumption into the optimized DL-LSTM-A deep network model to obtain quantiles;

[0048] The prediction module is used to process the quantiles using a non-parametric density estimation method to obtain the final load probability density function and complete the short-term load probability density prediction based on quantile regression.

[0049] In a third aspect, the present invention provides a short-term load probability density prediction system based on quantile regression, comprising a processor and a storage medium;

[0050] The storage medium is used to store instructions;

[0051] The processor is configured to operate according to the instructions to execute the steps of the method according to any one of the first aspects.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] The present invention proposes a double-layer LSTM cell structure to better extract high-order features of short-term power load forecasting, and introduces an attention mechanism to solve the key problems of each part of the long sequence problem, and obtains power load forecasting results at different quantiles. Finally, the Gaussian kernel function non-parametric estimation method is used to obtain the power load probability density prediction. The present invention can provide a more accurate long-term power load probability density prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments and in conjunction with the accompanying drawings, wherein:

[0055] Figure 1 A schematic diagram of a DL-LSTM-A cell structure according to an embodiment of the present invention;

[0056] Figure 2 A schematic diagram of the LSTM structure of each layer in an embodiment of the present invention;

[0057] Figure 3 A schematic diagram of a DL-LSTM-A network model according to an embodiment of the present invention;

[0058] Figure 4 This is a DL-LSTM-A quantile regression flow chart of an embodiment of the present invention;

[0059] Figure 5 A schematic diagram of the prediction performance of the probability density function of the DL-LSTM-A model according to an embodiment of the present invention;

[0060] Figure 6 Schematic diagram of median prediction performance under different models. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0062] The application principle of the present invention is described in detail below in conjunction with the accompanying drawings.

[0063] Example 1

[0064] An embodiment of the present invention provides a short-term load probability density prediction method based on quantile regression, comprising:

[0065] Obtain a DL-LSTM-A deep network model, wherein the DL-LSTM-A deep network model includes a plurality of double-layer LSTM network cells and an attention mechanism module, wherein the output ends of each double-layer LSTM network cell are respectively connected to the attention mechanism module; the attention mechanism module performs weighted summation on the outputs of each double-layer LSTM network cell; the double-layer LSTM network cell includes two LSTM cells connected in sequence, see Figure 3 ;

[0066] Use the training data to train and obtain the various parameters to be optimized in the DL-LSTM-A network, and obtain the optimized DL-LSTM-A deep network model;

[0067] The acquired influencing factors of power load consumption are input into the optimized DL-LSTM-A deep network model to obtain quantiles;

[0068] The non-parametric density estimation method is used to process the quantiles to obtain the final load probability density function and complete the short-term load probability density forecast based on quantile regression.

[0069] The method in the embodiment of the present invention is described in detail below in conjunction with a specific implementation manner.

[0070] First, the factors affecting power load consumption are analyzed. Through the analysis of the characteristics, the factors of power load consumption are mainly affected by the factors shown in Table I. The power load prediction value Y is modeled as follows:

[0071] Y=f(h,d,ty,s,te,n,l,Y o ) (1)

[0072] Where h∈{1,2,…,24} is the number of hours in a day, d∈{1,2,…,365} refers to a day in a year; ty is the type of day, ty=1 refers to a working day, ty=0 refers to a rest day; s refers to the season, {1,2,3,4} represents the four seasons of spring, summer, autumn and winter respectively; te refers to the temperature at the current time, described in degrees Celsius; n refers to the number of users served by the current terminal, Y o is the historical load value. f(·) is a nonlinear complex function used to describe the relationship between power load and various factors.

[0073] Table 1 Characteristic factors of power load

[0074]

[0075] The distribution function of the power load forecast value Y is defined as:

[0076] F(y)=P(Y≤y) (2)

[0077] Then the minimum value of y that satisfies F(y)≥τ is the τth quantile of variable y, which can be written as follows

[0078] Q(τ)=inf(y|F(y)≥τ) (3)

[0079] In the actual power grid, the relationship between power load and its influencing factors is nonlinear, and there is a correlation with the previous power load, so there is a memory characteristic. Based on the LSTM network, a double layer LSTM network with attention (DL-LSTM-A) deep network structure is proposed and introduced into the quantile regression analysis. Figure 1 The cell structure in the proposed DL-LSTM-A model is described. In each cell, two LSTM units are connected, and the output of the previous LSTM layer is used as the input of the next LSTM layer. This two-layer structure can better extract information from long sequences, thereby improving the performance of the model. In addition, the attention mechanism is introduced, such as Figure 1 As shown in the figure, each cell has an attention module, which inputs the current input and the current output through the tanh layer and obtains the attention value at that moment after normalization. The attention mechanism is used to analyze the importance of each cell, thereby further improving the effectiveness of the model.

[0080] For each LSTM layer in the DL-LSTM-attention network cell, it is composed of traditional LSTM cells, and its main structure is as follows: Figure 2 There are 4 gates in each LSTM layer, and the input-output relationship of each gate can be expressed by the following formula. The first one is called the forget gate, and the output of the first gate is f t and input h t-1 , X t And the parameter W f 、b f The relationship between can be calculated by the following relationship:

[0081] f t =σ(W f [h t-1 , X t ]+b f ) (4)

[0082] The second and third layers are called the input layer and tanh layer, respectively. The sigmoid function of the input layer determines which information to update. The tanh layer introduces a new candidate value. The relationship between the second and third layer outputs and the input and network parameters is described by the following formulas:

[0083] i t =σ(W i [h t-1 , X t ]+b i ) (5)

[0084]

[0085] At this time, the old state is updated to the new state. At this time, the output ht of the LSTM cell is obtained by the following formula

[0086]

[0087] O t =σ(W O [h t-1 , X t ]+b O ) t (8)

[0088] h t =O t tanh(C t ) (9)

[0089] Although the above LSTM network has memory characteristics, it cannot detect which is the important part of the sequence that affects the power load. To solve this problem, an attention mechanism is designed in the LSTM network to automatically utilize the importance of each part of the power load sequence at different times. Figure 3 The designed attention mechanism is described. In the attention mechanism of the DL-LSTM-A network, the output of the network is the weighted sum of each DL-LSTM-A network cell, which is expressed as follows:

[0090]

[0091] Where n+1 refers to the length of the cell sequence in the proposed attention mechanism, β k Refers to the attention value at the t-(k-1)th cell. Normalized using the norm, the attention value β k It can be calculated by the following formula

[0092]

[0093] Among them, ||·||2 refers to the second norm of the quantity. The s of each cell kUsed to describe the importance of cells. Figure 1 In order to better extract the characteristics of the load sequence, the present invention associates the attention value with the input, so for each moment s t Designed for

[0094]

[0095] in, and W a is a parameter that needs to be learned. From equations (11) and (12), we can see that the attention value at time t depends on the input X t and the cell output variable h at the current time t and its previous n time steps t The attention value can also be regarded as the activation function of the flow selection gate. t After passing through a fully connected layer, the final output is obtained. The above formulas describe the relationship between the output and input of each module in the DL-LSTM-A network and the parameters. According to the relationship between each module, the relationship between the output (i.e., power load prediction value) and input (characteristic factors) of the entire network and the network parameters is described as follows:

[0096] y=f DL-LSTM-A (X, W, V, b) (13)

[0097] After obtaining the above model, the parameters are still unknown. It is necessary to train the parameters according to the existing data, that is, to train the parameters (X, W, V, b) to make the output as accurate as possible. The classic Adam algorithm of deep network is used to optimize the parameters. In order to prevent the model from overfitting during the training process, regularization is considered when designing the cost function. The optimization objective function of training parameters W, V, b is:

[0098]

[0099] where ||·|| F refers to the matrix Frobenius norm, more precisely, the function f loss for:

[0100]

[0101] Among them, τ is the quantile, N is the number of samples, and Y i Refers to the output power sample, X i and X′ i All refer to samples under the quantile τ condition, i|Y i ≥f DL-LSTM-A (X i ′, W, V, b) indicates that the i-th response variable is greater than or equal to the regression estimate, i|Y i <f DL-LSTM-A (Xi ′, W, V, b) indicates that the i-th response variable is less than the regression estimate, the quantile τ∈(0,1), and the function ρ τ (u) = u(τ-I(u)), I(u) is called the characteristic function, u is the independent variable, expressed as:

[0102]

[0103] For the optimization problem of formula (18), the traditional optimization method can be used to solve it. The present invention uses the traditional adam algorithm to solve it and obtain the parameters W, V, and b. After the parameters are obtained, the quantile model of the power load can be obtained. From the above, it can be seen that the power load quantile can be obtained after the DL-LSTM-A model. At this time, the non-parametric density estimation method is used to process the quantile of the power load, and the final probability density function of the load can be obtained. Since the Gaussian kernel function non-parametric density estimation method has good mathematical characteristics and wide applicability, the present invention uses the Gaussian kernel function non-parametric density estimation method to obtain the probability density function of the power load. The specific process is as follows: the probability density function can be expressed by the Gaussian function as

[0104]

[0105] Where K(·) is the Gaussian kernel function, which is expressed as:

[0106]

[0107] Among them, σ is the smoothing parameter, also known as the bandwidth. Then the probability density function can be obtained by finding each smoothing parameter. The smoothing parameter has a great influence on the shape of the load density function. When the smoothing parameter bandwidth is too small, the points for calculating the system parameters are reduced, resulting in a decrease in the Gaussian kernel function deviation, an increase in variance, and an unstable estimation result. When the bandwidth is too large, the number of points used for calculation increases, the variance decreases, and the deviation may increase, but the method essentially requires the bandwidth to approach 0. Therefore, the choice of bandwidth needs to take into account the deviation-variance average. The choice of parameters has a great influence on the load probability density function, and the cross-validation criterion is used to obtain the parameters.

[0108] The main ideas and logical thinking are as follows Figure 4 shown.

[0109] The power data of a county in Anhui Province is used as a sample. It contains the power load data of multiple institutes and analyzes the power load data of each institute. The proposed DL-LSTM-A model is executed using the Tensorflow framework. Inteli7-8700 processor and Nvidia GeForce RTX 2080Ti graphics processor are used. When training the model, the batch size is 128 and the time series length is 48, that is, the data 48 hours in advance is used as a training data. In order to obtain training results faster, the data is first normalized and X is converted to t Normalized to

[0110]

[0111] The proposed DL-LSTM-A is compared with common probabilistic load forecasting algorithms such as support vector machine quantile regression and CNN-GRU quantile regression. The quantiles selected in all methods range from 0.01 to 0.99, which are 99 quantiles with an interval of 0.01. In order to better verify the effectiveness of the proposed scheme, the mean absolute percentage error (MAPE) and the maximum relative percentage error (MRPE) are used as indicators for comparison, which are calculated as follows:

[0112]

[0113]

[0114] The historical load data of Anhui Province in February and March 2021 is used as the basis. The adopted data set has undergone data preprocessing operations such as validity detection, missing value interpolation, and normalization to ensure data quality. In this data set, the power load data of every day and every hour is recorded. In February and March, the weather is usually cold. Using the historical load data from February 1 to March 20 as training data, the power load of 48 hours on March 21 and March 22 was predicted. The DS-LSTM-A model was used to predict the 48-hour prediction results on March 21 and March 22, as shown in the attached figure. Figure 5 As shown, attached Figure 5The probability density function, median value and actual power load value predicted by the model at four time points are given. These four time points are 5, 10, 15 and 20 hours. It can be seen from these four figures that the actual power load value is always above the probability density function of the proposed DL-LSTM-A model. At the 5th, 10th and 20th hours, the actual value of the power load is close to the value corresponding to the maximum probability density function predicted by the DL-LSTM-A model. At the 15th hour, there is a certain deviation between the load value corresponding to the maximum probability density function and the actual value, but it is still above the probability density line. In addition, by comparing the median value and the actual value predicted by the model in the four figures, it can be found that the median is close to the actual value. Therefore, it can be seen from the above analysis that the predicted value and prediction error obtained by the L-LSTM-A model can provide a better short-term power load forecasting method for power grid personnel, thereby providing reliable information for making power grid control strategies.

[0115] Attached Figure 6 The median predictions of the DL-LSTM-A, SVM, and CNN-GRU models are compared. Figure 6 The results show that compared with the SVR and CNN-GRU models, the median predicted by the DL-LSTM-A model is closer to the actual charge value, and the CNN-GRU has better performance than the SVM model. Since the DL-LSTM-A model uses a two-layer LSTM structure to better utilize the high-order features between charge sequences, the attention mechanism is introduced to distinguish the importance of each part of the sequence, thereby providing more accurate prediction results than other models, verifying the effectiveness of the DL-LSTM-A model.

[0116] In order to better describe the effectiveness of the proposed model, Table II describes the error statistics of the power load predicted by the DL-LSTM-A, SVM and CNN-GRU models, including MAPE and MRPE. As shown in Table 2, the average absolute error in the DL-LSTM-A model is 3.07%, and its maximum relative error is 4.58%, while the average absolute error of the CNN-GRU model is 4.66%, and its maximum relative error is 6.83%; the average relative error and maximum relative error of the SVM model are the largest, which are 6.09% and 9.46% respectively. It can be seen that DL-LSTM-A has the smallest relative error, can provide higher prediction accuracy, and has better prediction performance than other models.

[0117] Table 2 Statistics of prediction errors of different models

[0118] Model MAPE MRPE DL-LSTM-A 3.07％ 4.58％ SVM 6.09％ 9.46％ CNN-GRU 4.66％ 6.83％

[0119] Example 2

[0120] Based on the same inventive concept as that of Embodiment 1, an embodiment of the present invention provides a short-term load probability density prediction device based on quantile regression, comprising:

[0121] An acquisition module is used to acquire a DL-LSTM-A deep network model, wherein the DL-LSTM-A deep network model includes a plurality of double-layer LSTM network cells and an attention mechanism module, wherein the output ends of each double-layer LSTM network cell are respectively connected to the attention mechanism module; the attention mechanism module performs weighted summation on the outputs of each double-layer LSTM network cell; and the double-layer LSTM network cell includes two LSTM cells connected in sequence;

[0122] The optimization module is used to obtain various parameters to be optimized in the DL-LSTM-A network by using the training data to obtain the optimized DL-LSTM-A deep network model;

[0123] An execution module is used to input the acquired influencing factors of power load consumption into the optimized DL-LSTM-A deep network model to obtain quantiles;

[0124] The prediction module is used to process the quantiles using a non-parametric density estimation method to obtain the final load probability density function and complete the short-term load probability density prediction based on quantile regression.

[0125] Example 3

[0126] An embodiment of the present invention provides a short-term load probability density prediction system based on quantile regression, including a processor and a storage medium;

[0127] The storage medium is used to store instructions;

[0128] The processor is used to operate according to the instructions to execute the steps of the method according to any one of Embodiment 1.

[0129] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A short-term load probability density forecasting method based on quantile regression, characterized in that: include: Obtaining a DL-LSTM-A deep network model, wherein the DL-LSTM-A deep network model includes a plurality of double-layer LSTM network cells and an attention mechanism module, wherein the output ends of each double-layer LSTM network cell are respectively connected to the attention mechanism module; and the attention mechanism module performs weighted summation on the outputs of each double-layer LSTM network cell; The double-layer LSTM network cell includes two LSTM cells connected in sequence; Use the training data to train and obtain the various parameters to be optimized in the DL-LSTM-A network, and obtain the optimized DL-LSTM-A deep network model; The acquired influencing factors of power load consumption are input into the optimized DL-LSTM-A deep network model to obtain quantiles; The non-parametric density estimation method is used to process the quantiles to obtain the final load probability density function and complete the short-term load probability density forecast based on quantile regression. The method for obtaining the quantile includes: The calculation formula for the power load forecast value Y is defined as: Y=f(h,d,ty,s,te,n,l,Y o ) Where h∈{1,2,…,24} is the number of hours in a day, d∈{1,2,…,365} refers to a day in a year; ty is the type of day, ty=1 refers to a working day, ty=0 refers to a rest day; s refers to the season, {1,2,3,4} represents the four seasons of spring, summer, autumn and winter respectively; te refers to the temperature at the current time, described in degrees Celsius; n refers to the number of users served by the current terminal, Y o is the historical load value, f(·) is a nonlinear complex function used to describe the relationship between power load and various influencing factors; The distribution function of the power load forecast value Y is defined as: F(y)=P(Y≤y) Then the minimum y value that satisfies F(y)≥τ is the τth quantile of variable y, and the representation of the τth quantile is: Q(τ)=inf(y|F(y)≥τ).

2. A short-term load probability density forecasting method based on quantile regression according to claim 1, characterized in that: The output expression of the attention mechanism module is: Among them, n+1 refers to the length of the cell sequence in the proposed attention mechanism, β k refers to the attention value at the t-(k-1)th cell, h t-(k-1) Refers to the output value of the t-(k-1)th cell, H t Refers to the output of the attention mechanism, normalized using the norm, and the attention value β k It can be calculated using the following formula: Among them, ||·||2 refers to the second norm of the quantity, s k Used to describe the importance of cells, and then H t The output value obtained by inputting the fully connected layer is the output of the DL-LSTM-A network model, which is the predicted value of the power load.

3. The short-term load probability density forecasting method based on quantile regression according to claim 1 is characterized in that: Assume that after passing through the DL-LSTM-A network, the power load forecast value Y is related to the network parameters W, V, b and input X = [h, d, ty, s, te, n, l, Y o ] becomes Y=f DL-LSTM-A (X, W, V, b), then the parameters to be optimized in the DL-LSTM-A network obtained by training with the training data include W, V, and b, specifically: The design optimization objective function is: where ||·|| F refers to the matrix Frobenius norm, ||·||2 refers to the 2-norm, λ1, λ2, λ3 are three regularization variables, and the function f loss for: Among them, τ is the quantile, N is the number of samples, and Y i Refers to the output power sample, X i and X i ′ refers to the sample under the quantile τ condition, i|Y i ≥f DL-LSTM-A (X i ′,W,V,b) indicates that the i-th response variable is greater than or equal to the regression estimate, i|Y i <f DL-LSTM-A (X i ′,W,V,b) indicates that the i-th response variable is less than the regression estimate, the quantile τ∈(0,1), and the function ρ τ (u)=u(τ-I(u)), I(u) is called the indicator function, u is the independent variable, Based on the optimization objective function, the parameters are optimized in combination with the adam optimization algorithm to obtain the parameters to be optimized in the DL-LSTM-A network.

4. The short-term load probability density forecasting method based on quantile regression according to claim 1 is characterized in that: The probability density function is: in, is the Gaussian kernel function, x and x i is the independent variable, σ is the smoothing parameter, and the parameters in the Gaussian function are obtained using the cross-validation criterion.

5. The short-term load probability density forecasting method based on quantile regression according to claim 1 is characterized in that: Each LSTM cell consists of four gates. The first one is called the forget gate. The output of the first gate is f t and input h t-1 , X t And the first door parameter W f , b f The relationship between them is: f t =sigmoid(W f [h t-1 ,X t ]+b f ) sigmoid(·) is the sigmoid function; the second and third layers are called the input layer and tanh layer, respectively. The Tanh layer introduces a new candidate value. t and the second layer input h t-1 ,X t , the second layer parameter W i 、b i The relationship between them, and the third layer output With the third layer input h t-1 , X t and the third layer network parameter W C 、b C The relationship between them is described by the following formula: i t =sigmoid(W i [h t-1 ,X t ]+b i ) tanh(·) is the tanh function, with f t 、i t and the output C of the previous LSTM cell t-1 After that, the two outputs C of the LSTM cell t and h t It is obtained from the following formula: O t =sigmoid(W O [h t-1 ,X t ]+b O ) t h t =O t fishy(C) t ) Where W O 、b O is the parameter of the fourth layer, O t Intermediate variable, h t-1 , X t is the input parameter.

6. A short-term load probability density prediction device based on quantile regression, characterized in that: include: An acquisition module is used to acquire a DL-LSTM-A deep network model, wherein the DL-LSTM-A deep network model includes a plurality of double-layer LSTM network cells and an attention mechanism module, wherein the output ends of each double-layer LSTM network cell are respectively connected to the attention mechanism module; the attention mechanism module performs weighted summation on the outputs of each double-layer LSTM network cell; and the double-layer LSTM network cell includes two LSTM cells connected in sequence; The optimization module is used to obtain various parameters to be optimized in the DL-LSTM-A network by using the training data to obtain the optimized DL-LSTM-A deep network model; An execution module is used to input the acquired influencing factors of power load consumption into the optimized DL-LSTM-A deep network model to obtain quantiles; The prediction module is used to process the quantiles using a non-parametric density estimation method to obtain the final load probability density function and complete the short-term load probability density prediction based on quantile regression; The method for obtaining the quantile includes: The calculation formula for the power load forecast value Y is defined as: Y=f(h,d,ty,s,te,n,l,Y o ) Where h∈{1,2,…,24} is the number of hours in a day, d∈{1,2,…,365} refers to a day in a year; ty is the type of day, ty=1 refers to a working day, ty=0 refers to a rest day; s refers to the season, {1,2,3,4} represents the four seasons of spring, summer, autumn and winter respectively; te refers to the temperature at the current time, described in degrees Celsius; n refers to the number of users served by the current terminal, Y o is the historical load value, f(·) is a nonlinear complex function used to describe the relationship between power load and various influencing factors; The distribution function of the power load forecast value Y is defined as: F(y)=P(Y≤y) Then the minimum y value that satisfies F(y)≥τ is the τth quantile of variable y, and the representation of the τth quantile is: Q(τ)=inf(y|F(y)≥τ).

7. A short-term load probability density forecasting system based on quantile regression, characterized in that: including processors and storage media; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1-5.

Images