Short-Term Hierarchical Load Forecasting Method for Distribution Networks Based on CNN-LSTM Model and Quadratic Programming

Abstract

The present invention provides a short-term hierarchical load forecasting method for a distribution network based on a CNN-LSTM model and quadratic programming, including: obtaining the load and related data of different nodes in the distribution network and performing data preprocessing; respectively designing a prediction model input data set for each node; respectively establishing a CNN-LSTM model for load forecasting; and adjusting the above load forecasting results through quadratic programming to achieve short-term hierarchical load forecasting for the distribution network. On the premise of fully considering different influencing factors, different periodic characteristics of different nodes in the distribution system, and the hierarchical structure, the present invention establishes a CNN-LSTM prediction model with generalization ability for different nodes of the distribution network, and through the prediction post-processing technology of quadratic programming, the adjusted load meets the consistency requirements of the hierarchical structure between the loads of different-level nodes, realizing short-term hierarchical load forecasting for different nodes of the distribution network.

Description

Technical Field

[0001] The present invention relates to the short-term hierarchical load forecasting technology of a distribution network, and particularly to a short-term hierarchical load forecasting method for a distribution network based on a CNN (Convolutional Neural Networks)-LSTM (Long Short Term Memory network) model and quadratic programming. Background Art

[0002] The power industry is the lifeblood of urban production and life. In the distribution system, accurate short-term load forecasting for substations and transformers is of great significance for formulating power generation plans and optimizing the dispatching of the distribution network. In the distribution network, the load temperature sensitivity and periodic characteristics of different nodes are different, and the loads are significantly organized in a hierarchical structure among multiple nodes such as substations, transformers, and feeders. Therefore, a load forecasting model considering the different load characteristics of each node and a hierarchical load forecasting method considering the hierarchical structure of the distribution network are very important for the operation decision-making of the distribution system.

[0003] In recent years, deep learning intelligent algorithms have been increasingly used in short-term load forecasting because they can extract features during the training process and have stronger nonlinear fitting capabilities compared with traditional machine learning algorithms. Among them, the RNN network and its variant LSTM network are widely used in short-term load forecasting problems because of their time series forecasting capabilities, and the CNN model is widely used because of its strong feature extraction capabilities. Some complex network models that combine the CNN model and the LSTM model considering load periodicity and time dependence have been proposed to improve the forecasting accuracy, but most of these models ignore the influence of weather. In addition, most of the existing literature focuses on load forecasting for a single region, and directly forecasting the load of each level separately may make the forecasting results not meet the hierarchical structure requirements, that is, the sum of the bottom-layer forecasting results cannot be equal to the aggregated-layer forecasting result, which will have an adverse impact on subsequent decision-making. The existing hierarchical forecasting methods that consider the hierarchical requirements, such as bottom-up and top-down, first forecast the load of a certain level and then obtain the forecasting results of other levels through the hierarchical structure, and cannot fully utilize the load information of each layer. Summary of the Invention

[0004] The purpose of the present invention is to provide a short-term hierarchical load forecasting method for a distribution network based on a CNN-LSTM model and quadratic programming, design a CNN-LSTM model for different nodes in the distribution network to train and mine data for different daily type loads and corresponding temperature data, improve the short-term load forecasting accuracy of a single node, and design a quadratic programming model to adjust the load forecasting results of each node so that the final forecasting results meet the hierarchical structure requirements of the distribution network.

[0005] To achieve the above object, the present invention is implemented through the following technical solutions:

[0006] A short-term hierarchical load forecasting method for a distribution network based on a CNN-LSTM model and quadratic programming, comprising:

[0007] S1. Obtain the load data, local temperature data, and calendar information of different nodes in the distribution network, and preprocess the data;

[0008] S2. Divide the loads of different nodes in the distribution network into temperature-sensitive loads and temperature-insensitive loads, and design the input data sets of the prediction model for each node respectively;

[0009] S3. For different nodes in the distribution network, establish a CNN-LSTM prediction model based on the input data set, and perform future 24-hour load forecasting for the corresponding nodes based on each prediction model;

[0010] S4. Adjust the load forecasting results separately generated for different nodes in the distribution network through a prediction post-processing technique based on quadratic programming to obtain the final short-term hierarchical load forecasting results of different nodes in the distribution network.

[0011] Preferably, the preprocessing of the data in S1 includes:

[0012] Normalize each data value in the numerical data, where the numerical data includes load data and local temperature data; the normalization calculation formula is as follows:

[0013]

[0014] In the formula, represents the normalized value of the numerical data x, and x max and x min represent the maximum value and the minimum value in the numerical data respectively.

[0015] Preferably, the method for dividing the loads of different nodes in the distribution network into temperature-sensitive loads and temperature-insensitive loads in S2 includes:

[0016] Calculate the Pearson correlation coefficient between the daily average load and the daily average temperature of each node, and use the node load with the absolute value of the Pearson correlation coefficient greater than the preset threshold as the temperature-sensitive load, and the node load with the absolute value of the Pearson correlation coefficient not greater than the preset threshold as the temperature-insensitive load.

[0017] Preferably, the calculation formula of the Pearson correlation coefficient P xy is as follows:

[0018]

[0019] Wherein, g i , h i are the daily average load and daily average temperature corresponding to the i-th day, and are the average value of the daily average load and the average value of the daily average temperature.

[0020] Preferably, the input dataset of the prediction model in the S2 contains 7 time steps, and each time step contains the hourly load data, corresponding moment data, and corresponding week data of the 1-7 days before the day to be predicted. For temperature-sensitive loads, it also includes hourly temperature data.

[0021] Preferably, in the S3, the CNN-LSTM model includes a number of CNN modules, LSTM modules, and Dense modules. The input of each time step passes through one of the CNN modules to obtain a one-dimensional feature vector of that step, which is respectively used as the input of each time step of the LSTM module. Through the LSTM module and the Dense module, the load prediction value for the next 24 hours is obtained.

[0022] Preferably, the CNN module includes a convolutional layer and a max pooling layer. In the convolutional layer, the convolutional kernel size is 3*3, the convolutional step is 1*1, and the number of channels of the convolutional kernel is 32. The ReLU activation function is used. In the max pooling layer, the step is 1*1. For temperature-sensitive loads and temperature-insensitive loads, the filter sizes are 2*2 and 1*2 respectively.

[0023] Preferably, the LSTM network in the LSTM module is a many-to-one LSTM, and the hidden layer contains 200 LSTM units.

[0024] Preferably, the parameter update formula of the LSTM network in the LSTM module is:

[0025] f t =σ(W xf x t +W hg h t-1 +b f )

[0026] i t =σ(W xi x t +W hi h t-1 +b i )

[0027] g t =tanh(W xg x t +W hg h t-1 +bg )

[0028] c t = f t c t-1 + i t g t

[0029] o t = σ(W xo x t + W ho h t-1 + b o )

[0030] h t = o t tanh(c t )

[0031] wherein, x t is the unit input, h t is the unit output, c t is the unit state, f t is the forget gate, i t is the input gate, g t is the information for updating the unit state, o t is the output gate, W xf , W hf , W xi , W hi , W xg , W hg , W xo , W ho are weight matrices, b f , b i , b g , b o are offsets, σ represents the activation function, and tanh represents the hyperbolic tangent function.

[0032] Preferably, in the S4, for the hierarchical structure with one aggregation layer node and N bottom layer nodes in the distribution network, the method for adjusting the load forecasting results separately generated by different nodes in the distribution network through the prediction post-processing technology based on quadratic programming includes:

[0033] To meet the hierarchical structure requirements, that is, at all future times t, the load forecasting value of the aggregation layer node is equal to the sum of the load forecasting values of all bottom layer nodes, and at the same time, to minimize the forecasting value adjustment amount, a quadratic programming problem is constructed:

[0034]

[0035]

[0036] -Δ ≤ u ≤ Δ

[0037] wherein

[0038]

[0039]

[0040]

[0041] where the load prediction values of all nodes are the load prediction values of the nodes in the aggregation layer, are the load prediction values of the bottom - layer nodes, u ∈ R t×(N+1) is the adjustment variable, Q is the weight matrix of the adjustment variable u, and the coefficient A = [I t , -I t , …, -I t ∈ R t×(N+1) , I t is the t - order identity matrix, is the mean of the prediction errors in the validation set, M is the number of samples in the validation set, y i and represent the true load values and load prediction values of all nodes at each moment of the i - th sample in the validation set, Δ = [Δ1, …, Δ t×(N+1) ∈ R t×(N+1) is the upper limit of the adjustment variable u, and are respectively the true load value and load prediction value corresponding to the j - th element of the i - th sample in the validation set;

[0042] Adding the adjustment variable u to the original prediction result to obtain a load prediction result that meets the requirements of the distribution network hierarchy

[0043] Compared with the prior art, the present invention has the following advantages:

[0044] When performing load prediction on a single node in the distribution network, a CNN - LSTM prediction model is established. The model structure is designed by comprehensively considering the characteristics of load periodicity, time - dependence, and temperature - influence. Combining the advantages of the CNN model in feature extraction and the LSTM network model in time - series learning, the accuracy and precision of single - node load prediction are improved. By adjusting the individual load prediction results of each node through a prediction post - processing technology based on quadratic programming, the final load prediction result meets the requirements of the distribution network hierarchy and can be better applied to subsequent load scheduling decisions. By comparing with the load prediction results of other deep - learning methods and hierarchical prediction methods, the mean absolute percentage error (MAPE) of this method is lower, indicating that this method has a high prediction accuracy. Brief Description of the Drawings

[0045] To more clearly illustrate the technical solution of the present invention, the drawings required for description will be briefly introduced below. Obviously, the drawings in the following description are an embodiment of the present invention. For those of ordinary skill in the art, without creative work, other drawings can be obtained based on these drawings:

[0046] Figure 1 It is a flowchart of a short-term hierarchical load forecasting method for a distribution network based on a CNN-LSTM model and quadratic programming provided by an embodiment of the present invention;

[0047] Figure 2 It is a comparison chart of load forecasting curves for different deep learning models;

[0048] Figure 3 It is a comparison chart of load forecasting curves obtained by different hierarchical forecasting methods. Detailed Embodiment

[0049] The following further details the solution proposed by the present invention in conjunction with the drawings and specific embodiments. According to the following description, the advantages and features of the present invention will be clearer. It should be noted that the drawings are in a very simplified form and use non-precise scales, only for the purpose of facilitating and clearly assisting in explaining the embodiments of the present invention. In order to make the purpose, features, and advantages of the present invention more obvious and understandable, please refer to the drawings. It should be noted that the structures, scales, sizes, etc. shown in the drawings of this specification are only used to cooperate with the content disclosed in the specification for those skilled in this technology to understand and read, and are not used to limit the limiting conditions for the implementation of the present invention. Therefore, they do not have a technical essence. Any modification of the structure, change of the proportional relationship, or adjustment of the size, without affecting the effects that the present invention can produce and the purposes that can be achieved, should still fall within the scope covered by the technical content disclosed by the present invention.

[0050] The present invention comprehensively considers load forecasting models for different nodes of a distribution network applicable to different temperature sensitivities, different periodicities, and having long-term dependencies, and a prediction post-processing method that satisfies the hierarchical structure relationship between different nodes of the distribution network, and designs a short-term hierarchical load forecasting method for a distribution network based on a CNN-LSTM model and quadratic programming.

[0051] As Figure 1 shown, a short-term hierarchical load forecasting method for a distribution network based on a CNN-LSTM model and quadratic programming provided by the present invention specifically includes the following steps:

[0052] S1. Obtain load data, local temperature data, and calendar information of different nodes in the distribution network, and preprocess the data.

[0053] Specifically, load data of a certain substation and each transformer therein, daily temperature data of the corresponding area, week information data, and moment information data in the distribution network are obtained, and the data is preprocessed. Specifically, all numerical data is preprocessed by data normalization. Among them, the transformer load data and temperature data are numerical data. The normalization formula is as follows:

[0054]

[0055] In the formula, represents the normalized value of the numerical data x, and x max and x min represent the maximum value and the minimum value in the numerical data respectively.

[0056] S2. Classify the loads of different nodes in the distribution network into temperature-sensitive loads and temperature-insensitive loads, and design the input data sets of the prediction model for each node respectively.

[0057] First, the loads of different nodes in the distribution network can be classified into temperature-sensitive loads and temperature-insensitive loads. Specifically, the Pearson correlation coefficient between the daily average load and the daily average temperature of each node can be calculated, and classification can be performed according to the magnitude of the Pearson correlation coefficient. The node load with a Pearson correlation coefficient value greater than the preset threshold is used as the temperature-sensitive load, and the node load with a Pearson correlation coefficient value not greater than the preset threshold is used as the temperature-insensitive load.

[0058] The calculation formula of the Pearson correlation coefficient is as follows:

[0059]

[0060] In the formula, g i and h i are the daily average load and the daily average temperature corresponding to the i-th day, is the average value of the daily average load and the average value of the daily average temperature. The larger |P xy | is, the greater the degree of correlation. Taking the preset threshold as 0.4, the load with |P xy |>0.4 is the temperature-sensitive load, and the load with |P xy |≤0.4 is the temperature-insensitive load.

[0061] Generally speaking, the daily power load data is related to the power load data of the previous few days or the next few days, and the power load data of some nodes is also related to the power load data on the same day of the previous week. Therefore, the input of the CNN-LSTM prediction model includes the historical data of the seven days before the date to be predicted, that is, the input data set contains 7 time steps. The load at each moment of a day is also related to the corresponding moment, the corresponding week, and for temperature-sensitive loads, it is also related to the corresponding temperature. Therefore, in each time step, the input is a two-dimensional feature image composed of relevant feature data from 1 to 7 days before the date to be predicted. For temperature-sensitive loads, the feature image contains 7×24 pixel points, and each row is the load data, corresponding moment data, corresponding week data, corresponding maximum and minimum temperature data of the 24 hours of this day, as well as the corresponding maximum and minimum temperature data of the next day. For temperature-insensitive loads, the feature image contains 3×24 pixel points, and each row is the load data, corresponding moment data, and corresponding week data of the 24 hours of this day.

[0062] S3. For different nodes in the distribution network, based on the input data set, establish a CNN-LSTM prediction model respectively, and based on each of the prediction models, perform future 24-hour load prediction on the corresponding nodes;

[0063] In this embodiment, the CNN-LSTM model is composed of a CNN module, an LSTM module, and a Dense layer. The input of each time step passes through a CNN module to obtain a one-dimensional feature vector of this step. The outputs of the 7 CNN modules are respectively used as the inputs of each time step of the LSTM module. After passing through the LSTM module and through the Dense layer, the load prediction value for the next 24 hours is obtained.

[0064] Specifically, the CNN module is used to further extract features from the feature image, which includes a two-dimensional convolutional layer and a max pooling layer. In the convolutional layer, the kernel size of the convolutional kernel is 3*3, the convolutional stride is 1*1, and the number of channels of the convolutional kernel is 32. The ReLU activation function is used. In the max pooling layer, the stride is 1*1. For temperature-sensitive loads and temperature-insensitive loads, the filter sizes are 2*2 and 1*2 respectively. Each CNN module extracts a total of 352 features.

[0065] The LSTM module is a many-to-one LSTM network. The input is the extracted feature vectors of 7 time steps obtained by the CNN module, and the output is the load value for the next 24 hours. There are 200 LSTM nodes in the hidden layer. Finally, the load prediction value for the next 24 hours is obtained through the Dense layer with the ReLU activation function. The parameter update formula of the LSTM network in the LSTM module is:

[0066] f t =σ(W xf x t+W hf h t-1 +b f )

[0067] i t =σ(W xi x t +W hi h t-1 +b i )

[0068] g t =tanh(W xg x t +W hg h t-1 +b g )

[0069] c t =f t c t-1 +i t g t

[0070] o t =σ(W xo x t +W ho h t-1 +b o )

[0071] h t =o t tanh(c t )

[0072] Among them, x t is the unit input, h t is the unit output, c t is the unit state, f t is the forget gate, i t is the input gate, g t is the information used to update the unit state, o t is the output gate, W xf 、W hf 、W xi 、W hi 、W xg 、W hg 、W xo 、W ho are weight matrices, b f 、b i 、b g 、b o are offsets, σ represents the activation function, and tanh represents the hyperbolic tangent function.

[0073] During the training of the CNN-LSTM model, the Adam optimizer is selected, and the mean absolute error (MAE) is chosen as the loss function. The calculation formula of MAE is as follows:

[0074]

[0075] where N is the total number of prediction samples, p i is the load prediction value, and r i is the true load value.

[0076] S4. Through the prediction post-processing technology based on quadratic programming, the load prediction results separately generated by different nodes in the distribution network are adjusted to obtain the final short-term hierarchical load prediction results of different nodes in the distribution network.

[0077] In this embodiment, only the hierarchical structure composed of substations and N main transformers in the distribution network is considered. Then, there is 1 node in the aggregation layer and N nodes in the bottom layer. When predicting the load for the next t (t = 24 in the present invention) hours, the load prediction values of the substation and the N main transformers at the next t moments obtained in step S3 together constitute the load prediction value vector

[0078]

[0079] where, is the load prediction value of the aggregation layer, is the load prediction value of the bottom layer.

[0080] To meet the requirements of the hierarchical structure, that is, at all time points, the load prediction value of the aggregation layer node is equal to the sum of the load prediction values of all nodes in the bottom layer, and at the same time, the prediction value adjustment amount is minimized as much as possible. A quadratic programming problem is constructed:

[0081]

[0082]

[0083] -Δ ≤ u ≤ Δ

[0084] In the formula,

[0085]

[0086]

[0087]

[0088] where u ∈ R t×(N+1) is the adjustment variable. By adding the adjustment variable u to the original prediction result the load prediction result that meets the requirements of the distribution network hierarchical structure is obtained Q is the weight matrix of the adjustment variable, and the coefficient A = [I t , -I t , …, -I t ∈ R t×(N+1) , I t is the t-order identity matrix, is the mean of the prediction errors in the validation set, M is the number of samples in the validation set, y i and represent the true load values and load prediction values at each moment of all nodes of the i-th sample in the validation set. Δ = [Δ1, …, Δ t×(N+1) ∈ R t×(N+1) is the upper limit of the adjustment variable u, and are respectively the true load value and the load prediction value corresponding to the j-th element of the i-th sample in the validation set.

[0089] The adjustment variable u is obtained from the above quadratic programming problem, and the adjusted prediction result that meets the hierarchical structure requirements is obtained through

[0090] To verify the accuracy of the load prediction method in the present invention, load data (sampling time interval is 1 hour) of a substation and four main transformers in a certain area of the Shanghai power distribution network from July 1 to September 8, 2020 for a total of 100 days, daily maximum and minimum temperature data in the same area, as well as week information and time information are collected to form a data set. The data from July 1, 2020 to August 9, 2020 is taken as the training set, the data from August 10, 2020 to August 19, 2020 is taken as the validation set, and the data from August 20, 2020 to September 8, 2020 is taken as the test set. The mean absolute percentage error (MAPE) is selected to evaluate the short-term load prediction effect of each method. The smaller the MAPE, the higher the prediction accuracy. The calculation of MAPE is as follows:

[0091]

[0092] In the formula, N is the total number of prediction samples, p i is the load prediction value, and r i is the true load value.

[0093] The load of the substation (aggregation layer) and four main transformers (bottom layer) is predicted respectively using the MLP, LSTM network and CNN-LSTM model. The prediction errors are shown in Table 1, and the load prediction curve on August 22, 2020 is as Figure 2 shown.

[0094] Separate hierarchical load forecasting is carried out using separate load forecasting, bottom-up hierarchical load forecasting, and the hierarchical load forecasting method based on quadratic programming prediction post-processing technology in the present invention. When each node is predicted separately, the CNN-LSTM prediction model is used. The prediction errors are shown in Table 2, and the load forecasting curve on August 22, 2020 is as Figure 3 shown.

[0095] Table 1 Load forecasting errors of different deep learning models

[0096] MLP LSTM CNN-LSTM Aggregation layer 13.10％ 7.2％ 6.95％ Bottom layer 16.71％ 11.55％ 10.88％

[0097] Table 2 Load forecasting errors of different hierarchical load forecasting methods

[0098] Separate prediction Bottom-up Quadratic programming Aggregation layer 6.95％ 6.68％ 6.25％ Bottom layer 10.88％ 10.88％ 10.61％

[0099] From the prediction results, it can be seen that the CNN-LSTM prediction model used in the present invention has higher prediction accuracy in the bottom layer and the aggregation layer compared with the other two deep learning models. And the hierarchical prediction method based on the quadratic programming prediction post-processing technology shown in the present invention has better prediction effects in both levels of the hierarchical structure compared with the separate prediction and the traditional bottom-up hierarchical prediction method. It can be seen from Figure 3 this that the prediction post-processing method in the present invention adjusts the load prediction values of the aggregation layer and the bottom layer in a more accurate direction in most cases, especially when the prediction residuals are large. Therefore, the adjusted prediction accuracy is improved.

[0100] In summary, a short-term hierarchical load forecasting method for a distribution network based on a CNN-LSTM model and quadratic programming provided by the above embodiments of the present invention includes the following steps: obtaining the load and related data of different nodes in the distribution network and performing data preprocessing; respectively designing the prediction model input data sets for each node; respectively establishing a CNN-LSTM model for load forecasting; performing prediction post-processing on the load forecasting results of all nodes in the distribution network through quadratic programming to obtain the final short-term hierarchical load forecasting results of different nodes in the distribution network. The present invention fully considers different influencing factors, different cycle characteristics of different nodes in the distribution system, and the hierarchical structure, establishes a CNN-LSTM prediction model with generalization ability for different nodes in the distribution network, and makes the adjusted load meet the consistency requirements of the hierarchical structure between the loads of different-level nodes through the prediction post-processing technology of quadratic programming, realizing the short-term hierarchical load forecasting of different nodes in the distribution network.

[0101] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the said element.

[0102] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be recognized that the above description should not be considered as a limitation of the present invention. After those skilled in the art have read the above content, various modifications and alternatives to the present invention will be obvious. Therefore, the protection scope of the present invention should be defined by the appended claims.

Claims

1. A short-term hierarchical load forecasting method for distribution networks based on the CNN-LSTM model and quadratic programming, characterized in that Including: S1. Obtain the load data, local temperature data, and calendar information of different nodes in the distribution network, and preprocess the data; S2. Divide the loads of different nodes in the distribution network into temperature-sensitive loads and temperature-insensitive loads, and design the input data sets of the prediction model for each node respectively; S3. For different nodes in the distribution network, establish a CNN-LSTM prediction model based on the input data set, and perform load prediction for the corresponding nodes in the next 24 hours based on each prediction model; S4. Adjust the load prediction results separately generated for different nodes in the distribution network through a prediction post-processing technique based on quadratic programming to obtain the final short-term hierarchical load prediction results of different nodes in the distribution network; The method of dividing the loads of different nodes in the distribution network into temperature-sensitive loads and temperature-insensitive loads in S2 includes: Calculate the Pearson correlation coefficient between the daily average load and the daily average temperature of each node, and regard the node load with the absolute value of the Pearson correlation coefficient greater than the preset threshold as the temperature-sensitive load, and the node load with the absolute value of the Pearson correlation coefficient not greater than the preset threshold as the temperature-insensitive load; The calculation formula of the Pearson correlation coefficient Pxy is as follows: where \(g_i\) and \(h_i\) are the daily average load and daily average temperature corresponding to the \(i\)-th day, and are the mean values of the daily average load and the daily average temperature; In S3, the CNN-LSTM model includes several CNN modules, LSTM modules, and Dense layers. The input of each time step passes through a CNN module to obtain the one-dimensional feature vector of this step, and is respectively used as the input of each time step of the LSTM module. The load prediction value for the next 24 hours is obtained through the LSTM module and the Dense layer; In S4, for a hierarchical structure in the distribution network with one aggregation layer node and N bottom layer nodes, the method of adjusting the load prediction results separately generated for different nodes in the distribution network through a prediction post-processing technique based on quadratic programming includes: To meet the hierarchical structure requirements, that is, at all future times t, the load prediction value of the aggregation layer node is equal to the sum of the load prediction values of all bottom layer nodes, and at the same time, minimize the prediction value adjustment amount, construct a quadratic programming problem: -Δ≤u≤Δ In the formula, Among them, the load prediction values of all nodes are the load prediction values of the nodes in the aggregation layer, are the load prediction values of the nodes in the bottom layer, u ∈ Rt×(N+1) is the adjustment variable, Q is the weight matrix of the adjustment variable u, the coefficient A = [It, -It,..., -It] ∈ Rt×(N+1), and It is the t-order identity matrix. is the mean of the prediction errors in the validation set, M is the number of samples in the validation set, yi and represent the true load values and load prediction values of all nodes at each moment of the i-th sample in the validation set. Δ = [Δ1,..., Δt×(N+1)] ∈ Rt×(N+1) is the upper limit of the adjustment variable u. and are the true load value and load prediction value corresponding to the j-th element of the i-th sample in the validation set, respectively. On the original prediction results Add the adjustment variable u to obtain the load prediction results that meet the requirements of the distribution network hierarchy 2. The short-term hierarchical load forecasting method for a distribution network based on a CNN-LSTM model and quadratic programming according to claim 1, characterized in that The preprocessing of the data in S1 includes: Normalize each data value in the numerical data. The numerical data includes load data and local temperature data. The normalization calculation formula is as follows: In the formula, represents the normalized value of the numerical data x, and xmax and xmin respectively represent the maximum value and the minimum value in the numerical data.

3. The short-term hierarchical load forecasting method for a distribution network based on a CNN-LSTM model and quadratic programming according to claim 1, characterized in that The input data set of the prediction model in S2 contains 7 time steps. Each time step contains the hourly load data, corresponding moment data, and corresponding week data of 1-7 days before the day to be predicted. For temperature-sensitive loads, it also includes hourly temperature data.

4. The short-term hierarchical load forecasting method for a distribution network based on a CNN-LSTM model and quadratic programming according to claim 1, characterized in that The CNN module includes a convolutional layer and a max pooling layer. The convolutional kernel size in the convolutional layer is 3*3, the convolutional step is 1*1, the number of channels of the convolutional kernel is 32, and the ReLU activation function is used. The step in the max pooling layer is 1*1. For temperature-sensitive loads and temperature-insensitive loads, the filter sizes are 2*2 and 1*2 respectively.

5. The short-term hierarchical load forecasting method for a distribution network based on a CNN-LSTM model and quadratic programming according to claim 1, characterized in that In the LSTM module, the LSTM network is a many-to-one LSTM, and the hidden layer contains 200 LSTM units.

6. The short-term hierarchical load forecasting method for a distribution network based on a CNN-LSTM model and quadratic programming according to claim 5, wherein The parameter update formula of the LSTM network in the LSTM module is as follows: ft = σ(Wxfxt + Whfht-1 + bf) it = σ(Wxixt + Whiht-1 + bi) gt = tanh(Wx9xt + Whght-1 + bg) ct = ftct-1 + itgt ot = σ(Wxoxt + Whoht-1 + bo) ht = ottanh(ct) Among them, xt is the unit input, ht is the unit output, ct is the unit state, ft is the forget gate, it is the input gate, gt is the information used to update the unit state, ot is the output gate, Wxf, Whf, Wxi, Whi, Wxg, Whg, Wxo, Who are weight matrices, bf, bi, bg, bo are offsets, σ represents the activation function, and tanh represents the hyperbolic tangent function.

Images