Substation multi-level power load prediction method and system

Abstract

The invention discloses a transformer substation multi-level power load prediction method and system, and the method comprises the steps: obtaining the transformer substation load element data of each type of transformer substations based on a clustering result obtained through the optimization clustering of all transformer substations in advance, and obtaining the economic data of a first-level region and the meteorological data of a second-level region; load prediction is carried out based on the first neural network model, and load prediction results of all types of substations are given; load prediction is carried out based on a second neural network model, and load prediction results of all second-level regions are given; and carrying out load prediction based on the third neural network model, and giving a load prediction result of the first-level region. According to the method, massive load data can be quickly and effectively processed.

Description

Technical Field

[0001] The present invention belongs to the technical field of power load prediction, and in particular relates to a multi-level power load prediction method and system for a substation. Background Art

[0002] Electricity load forecasting is the basis for the power system to formulate development plans and power generation plans. Improving the accuracy of electricity load forecasting is of great significance for the rational dispatch of power grid resources and the efficient and stable operation of new power systems. With the construction of new power systems, the high proportion of new energy access and its inherent random volatility make the load changes more complex. Electricity load is affected by many factors such as policy macro-control, economy, and meteorology. These factors are random and changeable. The nonlinear relationship between electricity load and influencing factors is complex. It is difficult to express the relationship between the two with an accurate relationship or to achieve prediction with a conventional prediction model. For example, patent CN110796293A discloses a method for electricity load forecasting, which involves the field of power technology. This method uses a trained least squares support vector machine model for longitudinal prediction for complete power data, and then uses a trained cubature Kalman filter prediction model for horizontal prediction. Finally, a trained gray neural network model is used to fuse the two algorithms to obtain the final prediction result; therefore, a new and more accurate load forecasting method is needed. With the development of smart grids, massive load data at all levels of the power system have been accumulated on the terminal side. How to utilize massive load data, fully consider the factors affecting load changes, and adopt effective load forecasting models to improve load forecasting accuracy is the key to load forecasting. Summary of the invention

[0003] In view of the defects in the above-mentioned prior art, the present invention provides a multi-level power load forecasting method and system for substations. Based on the multi-level relationship of loads, and taking into account load elements, socio-economic factors and meteorological factors, massive load data can be processed quickly and effectively.

[0004] In a first aspect, the present invention provides a multi-level power load forecasting method for a substation, comprising:

[0005] Based on the clustering results of pre-optimized clustering of all substations, substation load element data of each category of substations are obtained, and first-level regional economic data and second-level regional meteorological data are obtained; wherein the first-level region includes multiple second-level regions, and the second-level region includes multiple substations;

[0006] Based on the pre-built first neural network model, the substation load factor data of each type of substation and the meteorological data of the area corresponding to the substation are input to perform load forecasting, and the load forecast results of all types of substations are given;

[0007] Based on the pre-built second neural network model, the load forecast results of all types of substations, the second-level regional load factor data and the second-level regional meteorological data are input to perform load forecasting, and the load forecast results of all second-level regions are given;

[0008] Based on the pre-built third neural network model, the load forecast results of all second-level regions, the load factor data of the first-level regions and the economic data of the first-level regions are input for load forecasting, and the load forecast results of the first-level regions are given.

[0009] Furthermore, the substation load element data includes the substation load historical data, the substation load ratio and load rate, the substation line structure, and the corresponding new energy output of the substation. The second-level regional load element data includes the second-level regional load historical data, the second-level regional industrial structure, and the second-level regional new energy output. The first-level regional load element data includes the first-level regional load historical data, the first-level regional industrial structure, and the first-level regional new energy output.

[0010] Furthermore, when the first neural network model, the second neural network model and the third neural network model are pre-built, it also includes:

[0011] The time series relationship of multi-level power loads is established, and the first neural network model, the second neural network model and the third neural network are modified.

[0012] Furthermore, the time series relationship of the multi-level power load includes the time series relationship between the load data of the substation and the load data of the second-level region, and the time series relationship between the load history data of the second-level region and the load history data of the first-level region, which all satisfy the following relationship:

[0013]

[0014] In the formula, y is the upper load data, x1, x2, ..., x n is the n lower-level load data corresponding to the upper level, n is the number of lower-level loads, a i is the correction coefficient corresponding to the i-th lower-level load data, and c is a constant.

[0015] Further, the first neural network model, the second neural network model and the third neural network are modified, including:

[0016] Based on the time series relationship between the load data of the substation and the historical load data of the second-level area, the actual load result of the second-level area is determined by the actual load result of the substation;

[0017] Based on the time series relationship between the load history data of the second-level region and the load history data of the first-level region, the actual load result of the first-level region is determined by the actual load result of the second-level region;

[0018] The actual load results of the substation, the actual load results of the second-level area, and the actual load results of the first-level area are applied to the training of the first neural network model, the second neural network model, and the third neural network.

[0019] Furthermore, all substations are optimized and clustered in advance, including:

[0020] Resampling the data set containing the historical load curves of all substations in the first-level area to obtain multiple new data sets;

[0021] Cluster each new data set and give the cluster center corresponding to each new data set;

[0022] All cluster centers are combined to form a central data set, and hierarchical clustering is performed to obtain multiple clusters;

[0023] Traverse the original data set containing the historical load curves of all substations in the first-level area and the central data set, and match each historical load curve with the cluster center closest to it from the central data set;

[0024] Based on the matching results, the historical load curves in the original data set are divided into clusters corresponding to the matching cluster centers;

[0025] Based on the clusters divided by historical load curves and their corresponding relationships with substations, the results of substation division into different categories are obtained.

[0026] Furthermore, the pre-construction of the third neural network model includes:

[0027] Obtain historical first-level regional data to form a first-level regional data training set; the first-level regional data includes actual load results of all second-level regions, first-level regional load element data, first-level regional economic data and actual load results of the first-level regions;

[0028] The long short-term memory network is trained based on the first-level regional data training set until convergence, and the parameters to be learned in the long short-term memory network are updated to obtain the third neural network model.

[0029] Furthermore, the architectures of the first neural network model and the second neural network model are both optimized long short-term memory networks formed by adding two Sigmoid gating units to the long short-term memory network;

[0030] The pre-construction of the first neural network model and the second neural network model includes:

[0031] Obtain historical substation-level data and second-level regional data to form a substation-level data training set and a second-level regional data training set respectively; wherein the substation-level data includes substation load element data of each category of substation, meteorological data of the area corresponding to the substation, and actual load results of all categories of substations, and the second-level regional data includes actual load results of all categories of substations, second-level regional load element data, second-level regional meteorological data, and actual load results of all second-level regions;

[0032] Based on the substation-level data training set and the second-level regional data training set, the optimized long short-term memory network is trained until convergence, the parameters to be learned in the optimized long short-term memory network are updated, and the first neural network model and the second neural network model are obtained respectively.

[0033] Furthermore, two Sigmoid gate units are added to the long short-term memory network, including:

[0034] On the basis of the long short-term memory network architecture, the hidden unit at the previous moment is combined with a Sigmoid gating unit to obtain the optimized hidden unit at the previous moment, the input of the long short-term memory network is combined with another Sigmoid gating unit to obtain the optimized input of the long short-term memory network, and the optimized hidden unit at the previous moment and the optimized input of the long short-term memory network are respectively subjected to linear matrix transformation and combined to obtain the cell state; wherein, the Sigmoid gating unit combined with the hidden unit at the previous moment is connected to the input of the long short-term memory network, and the Sigmoid gating unit combined with the input of the long short-term memory network is connected to the optimized hidden unit at the previous moment.

[0035] Furthermore, the long short-term memory network is optimized to satisfy the following relationship:

[0036] u t =σ(W u x t +b u )

[0037] v t =σ(W v h t-1 +b v )

[0038] h′ t-1 =2h t-1 ⊙u t

[0039] x′ t =2x t ⊙vt

[0040] f t =σ(W hf h′ t-1 +W xf x′ t +b f )

[0041] i t =σ(W hi h′ t-1 +W xi x′ t +b i )

[0042]

[0043] o t =σ(W ho h′ t-1 +W xo x′ t +b o )

[0044] h t =o t ⊙tanh(c t )

[0045] In the formula, u t and v t are the Sigmoid gated unit combined with the hidden unit of the previous moment and the Sigmoid gated unit combined with the input of the long short-term memory network, respectively. u and w v are the linear weight matrices for linear matrix transformation of the optimized hidden unit and the optimized input of the long short-term memory network at the previous moment, and b u and b v are the linear bias matrices when the linear matrix transformation is performed on the optimized hidden unit of the previous moment and the optimized input of the long short-term memory network, x t and x' t are the input and optimization input of the long short-term memory network, respectively, h t-1 and h' t-1 are the hidden units at the previous moment and the optimized hidden units, respectively, t 、i t , o t are the output values ​​of the forget gate, input gate, and output gate respectively. is the memory unit before updating, c t is the updated memory unit, c t-1 is the memory unit updated at the last moment, [W hf Wxf ]、[W hi W xi ]、[W hg W xg ]、[W ho W xo ] are the network weights of the forget gate, input gate, memory unit before update and output gate, respectively. f , b i , b g , b o are the network biases of the forget gate, input gate, memory unit before update, and output gate, σ() is the Sigmoid logic activation function, ⊙ is the Hadamard product operator of the matrix, and h t is the hidden unit at the current moment.

[0046] Furthermore, when performing load forecasting, the new energy output corresponding to the substation and the meteorological data of the area corresponding to the substation, and the new energy output of the second-level area and the meteorological data of the second-level area are respectively used as the first neural network model and the second neural network model after forming correlation coefficients;

[0047] The following relationships are satisfied between the renewable energy output corresponding to the substation and the meteorological data of the area corresponding to the substation, as well as between the renewable energy output of the second-level area and the meteorological data of the second-level area of ​​the city:

[0048]

[0049] Among them, r is the correlation coefficient between energy output and meteorological data, x i ,y i They are the new energy output value and meteorological data value within a limited time period. are the average values ​​of the new energy output value and the meteorological data value within the limited time period, and n is the number of new energy output values ​​or meteorological data values ​​within the limited time period.

[0050] In a second aspect, the present invention further provides a substation multi-level power load forecasting system, which adopts the above-mentioned substation multi-level power load forecasting method, and the system includes:

[0051] A data acquisition module, which is used to acquire the second-level regional meteorological data and the first-level regional economic data, and acquire the substation load element data of each category of substation based on the clustering result of pre-optimized clustering of all substations;

[0052] A substation load prediction module is used to perform load prediction based on a pre-built first neural network model by inputting substation load element data of each type of substation and meteorological data of the area corresponding to the substation, and providing load prediction results for all types of substations;

[0053] The second-level regional load forecasting module is used to perform load forecasting based on the pre-built second neural network model, input the load forecasting results of all types of substations, the second-level regional load factor data and the second-level regional meteorological data, and provide the load forecasting results of all second-level regions;

[0054] The first-level regional load forecasting module is used to perform load forecasting based on the pre-built third neural network model, input the load forecasting results of all second-level regions, the first-level regional load factor data and the first-level regional economic data, and give the load forecasting results of the first-level regions.

[0055] Furthermore, the substation multi-level power load forecasting system also includes a data processing module, which includes:

[0056] The time series relationship of multi-level power loads is established, and the first neural network model, the second neural network model and the third neural network are modified.

[0057] Furthermore, the data processing module also includes:

[0058] Based on the time series relationship between the load data of the substation and the historical load data of the second-level area, the actual load result of the second-level area is determined by the actual load result of the substation;

[0059] Based on the time series relationship between the load history data of the second-level region and the load history data of the first-level region, the actual load result of the first-level region is determined by the actual load result of the second-level region;

[0060] The actual load results of the substation, the actual load results of the second-level area, and the actual load results of the first-level area are applied to the training of the first neural network model, the second neural network model, and the third neural network.

[0061] Furthermore, the data acquisition module includes:

[0062] Resampling the data set containing the historical load curves of all substations in the first-level area to obtain multiple new data sets;

[0063] Cluster each new data set and give the cluster center corresponding to each new data set;

[0064] All cluster centers are combined to form a central data set, and hierarchical clustering is performed to obtain multiple clusters;

[0065] Traverse the original data set containing the historical load curves of all substations in the first-level area and the central data set, and match each historical load curve with the cluster center closest to it from the central data set;

[0066] Based on the matching results, the historical load curves in the original data set are divided into clusters corresponding to the matching cluster centers;

[0067] Based on the clusters divided by historical load curves and their corresponding relationships with substations, the results of substation division into different categories are obtained.

[0068] Furthermore, the substation multi-level power load forecasting system also includes a model building module, which includes:

[0069] Obtain historical first-level regional data to form a first-level regional data training set; the first-level regional data includes actual load results of all second-level regions, first-level regional load element data, first-level regional economic data and actual load results of the first-level regions;

[0070] The long short-term memory network is trained based on the first-level regional data training set until convergence, and the parameters to be learned in the long short-term memory network are updated to obtain the third neural network model.

[0071] Furthermore, the model building module includes:

[0072] Obtain historical substation-level data and second-level regional data to form a substation-level data training set and a second-level regional data training set respectively; wherein the substation-level data includes substation load element data of each category of substation, meteorological data of the area corresponding to the substation, and actual load results of all categories of substations, and the second-level regional data includes actual load results of all categories of substations, second-level regional load element data, second-level regional meteorological data, and actual load results of all second-level regions;

[0073] Based on the substation-level data training set and the second-level regional data training set, the optimized long short-term memory network is trained until convergence, the parameters to be learned in the optimized long short-term memory network are updated, and the first neural network model and the second neural network model are obtained respectively.

[0074] Furthermore, the model building module also includes:

[0075] On the basis of the long short-term memory network architecture, the hidden unit at the previous moment is combined with a Sigmoid gating unit to obtain the optimized hidden unit at the previous moment, the input of the long short-term memory network is combined with another Sigmoid gating unit to obtain the optimized input of the long short-term memory network, and the optimized hidden unit at the previous moment and the optimized input of the long short-term memory network are respectively subjected to linear matrix transformation and combined to obtain the cell state; wherein, the Sigmoid gating unit combined with the hidden unit at the previous moment is connected to the input of the long short-term memory network, and the Sigmoid gating unit combined with the input of the long short-term memory network is connected to the optimized hidden unit at the previous moment.

[0076] The present invention provides a method and system for predicting multi-level power load in a substation, which has at least the following beneficial effects:

[0077] (1) Based on the first-level, second-level, and substation-level load data, meteorological factors and socio-economic factors are considered, and the load elements at each level are analyzed. An improved clustering algorithm and a three-layer stacked improved long short-term memory neural network algorithm prediction model are used to complete the bottom-up prediction of multi-level loads. This can significantly improve the prediction accuracy of multi-level power loads and has application value for load forecasting.

[0078] (2) Clustering by using an improved clustering algorithm (optimized clustering algorithm) can effectively avoid the problem that the K-Means clustering algorithm is easily affected by the initial K value and easily falls into the local optimal solution.

[0079] (3) By improving the long short-term memory network, the dynamic prediction ability of the constructed model can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 A flowchart of a multi-level power load forecasting method for a substation provided by the present invention;

[0081] Figure 2 A flow chart of clustering power loads provided in an embodiment of the present invention;

[0082] Figure 3 A schematic diagram of an optimized long short-term memory network provided by an embodiment of the present invention;

[0083] Figure 4 A curve diagram of first-level substation load clustering provided by an embodiment of the present invention;

[0084] Figure 5 A schematic diagram of the relationship between summer photovoltaic output and weather conditions provided by an embodiment of the present invention;

[0085] Figure 6 A schematic diagram showing a comparison of average load forecast values ​​for July in City Y provided by an embodiment of the present invention;

[0086] Figure 7 A schematic diagram showing a comparison of average load forecast values ​​for July in Province S provided by an embodiment of the present invention;

[0087] Figure 8 A schematic diagram of a substation multi-level power load forecasting system provided by the present invention. DETAILED DESCRIPTION

[0088] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0089] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms "a", "said" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings, and "multiple" generally includes at least two.

[0090] It should also be noted that the term "includes", "comprising" or any other variation thereof is intended to cover non-exclusive inclusion, so that a commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprising a ..." do not exclude the existence of other identical elements in the commodity or device including the elements.

[0091] like Figure 1 As shown, the present invention provides a multi-level power load forecasting method for a substation, comprising:

[0092] Based on the clustering results of pre-optimized clustering of all substations, substation load element data of each category of substations are obtained, and first-level regional economic data and second-level regional meteorological data are obtained; wherein the first-level region includes multiple second-level regions, and the second-level region includes multiple substations;

[0093] Based on the pre-built first neural network model, the substation load factor data of each type of substation and the meteorological data of the area corresponding to the substation are input to perform load forecasting, and the load forecast results of all types of substations are given;

[0094] Based on the pre-built second neural network model, the load forecast results of all types of substations, the second-level regional load factor data and the second-level regional meteorological data are input to perform load forecasting, and the load forecast results of all second-level regions are given;

[0095] Based on the pre-built third neural network model, the load forecast results of all second-level regions, the load factor data of the first-level regions and the economic data of the first-level regions are input for load forecasting, and the load forecast results of the first-level regions are given.

[0096] The first-level region can be the region corresponding to the provincial level, and the second-level region can be the region corresponding to the municipal level. The present invention takes the provincial-municipal-substation level as an example to illustrate the multi-level power load prediction method of the substation. Specifically, the substation load element data may include the substation load historical data, the load capacity ratio and load rate of the substation, the substation line structure, and the new energy output corresponding to the substation; the municipal load element data (corresponding to the second-level regional load element data) includes the load historical data of all districts and counties in the city (corresponding to the load historical data of the second-level region), the municipal industrial structure (corresponding to the second-level regional industrial structure), and the new energy output in the city (corresponding to the new energy output in the second-level region); the provincial load element data (corresponding to the first-level regional load element data) includes the load historical data of all prefecture-level cities in the province (corresponding to the load historical data of the first-level region), the provincial industrial structure (corresponding to the first-level regional industrial structure), and the new energy output in the province (corresponding to the new energy output in the first-level region). Among them, the substation line structure includes the 220kV substation line structure and the 110kV substation line structure. The 220kV substation line structure includes its subordinate 110kV substation and 110kV outgoing line; the 110kV substation line structure includes its subordinate 35kV substation and 10kV outgoing line. The industrial structure includes the proportion of the primary industry, secondary industry, and tertiary industry stipulated by the state and the economic development situation.

[0097] When the first neural network model, the second neural network model and the third neural network model are pre-built, it also includes:

[0098] The time series relationship of multi-level power loads is established, and the first neural network model, the second neural network model and the third neural network are modified.

[0099] The time series relationship of multi-level power loads includes the time series relationship between substation load data and the load data of all districts and counties in the city, and the time series relationship between the load history data of all districts and counties in the city and the load history data of all prefecture-level cities in the province, all of which satisfy the following relationship:

[0100]

[0101] Where y is the historical data of the upper load, x1, x2, ..., x n is the historical data of n lower-level loads corresponding to the upper level, n is the number of lower-level loads, a i is the correction coefficient corresponding to the i-th lower-level load historical data, and c is a constant.

[0102] The first neural network model, the second neural network model and the third neural network are modified, including:

[0103] Based on the time series relationship between the load data of the substation and the load data of all districts and counties in the city, the actual load results at the city level are determined through the actual load results of the substation;

[0104] Based on the time series relationship between the load history data of all districts and counties in the city and the load history data of all prefecture-level cities in the province, the actual load results at the provincial level are determined through the actual load results at the municipal level;

[0105] The actual load results of the substation, the actual load results of the municipal level and the actual load results of the provincial level are applied to the training of the first neural network model, the second neural network model and the third neural network.

[0106] like Figure 2 As shown, the power loads of all substations are clustered based on the optimized clustering algorithm, and the substation load element data of each category of substations are given, including:

[0107] Resample the data set containing the historical load curves of all substations in the province to obtain multiple new data sets;

[0108] Cluster each new data set and give the cluster center corresponding to each new data set;

[0109] All cluster centers are combined to form a central data set, and hierarchical clustering is performed to obtain multiple clusters;

[0110] Traverse the original data set containing the historical load curves of all substations in the province and the central data set, and match each historical load curve with the cluster center closest to it from the central data set;

[0111] Based on the matching results, the historical load curves in the original data set are divided into clusters corresponding to the matching cluster centers;

[0112] Based on the clusters divided by historical load curves and their corresponding relationships with substations, the results of substation division into different categories are obtained.

[0113] In actual application scenarios, the power loads of all substations are clustered based on the optimized clustering algorithm, that is, the power loads of all substations are clustered based on the K-Means clustering combined with the hierarchical clustering algorithm to avoid the problem that the K-Means clustering algorithm is easily affected by the initial K value and easily falls into the local optimal solution. Specifically, clustering can include the following steps:

[0114] (1) For a data set X containing N load curves (load curves of N substations at the substation level) N Perform bootstrap resampling to obtain multiple new load samples Among them, each sample is the data sample obtained by resampling the load curves of N substations, S is the number of samplings, specifically, there are N load curves before resampling, and N*S curves after resampling;

[0115] (2) Perform K-Means clustering on each new load sample to obtain S×M cluster centers c 11 ,c 12 ,…,c 1S ,c 21 ,…,c SM , where M is the number of clusters for each new load sample, c ij is the jth cluster center of the i-th new load sample;

[0116] (3) Combine all cluster centers into a new data set C S ={c 11 ,c 12 ,…,c SM}, for data set C S Perform hierarchical clustering to obtain K clusters

[0117] (4) For the original load curve data set X N For any load curve object x, if c(x)∈C S represents the cluster center closest to x, then x is divided into the cluster containing c(x); x and c(x) are both general terms, which are any load curve object and cluster center;

[0118] (5) All objects x∈X in the original load curve data N The clustering is completed when the data is divided into K clusters.

[0119] The present invention first expands the data set by resampling, obtains K clusters by hierarchical clustering, and then divides the original load curve into the K clusters, which can effectively avoid the influence of the initial K value and the problem of being easily trapped in the local optimal solution.

[0120] When load forecasting is performed in sequence through the pre-built first neural network model, the second neural network model and the third neural network model, the first neural network model and the second neural network model are improvements made on the basis of the third neural network model. Specifically, the first neural network model and the second neural network model both add two Sigmoid gating units to the third neural network model.

[0121] The pre-construction of the third neural network model may include:

[0122] Obtain historical provincial data to form a provincial data training set; provincial data includes all city-level actual load results, provincial load factor data, provincial economic data and provincial actual load results;

[0123] Based on the provincial data training set, the long short-term memory network (LSTM) is trained until convergence, and the parameters to be learned in the long short-term memory network are updated to obtain the third neural network model.

[0124] Among them, the convergence condition is that the error between the provincial load forecast result and the provincial actual load result is within the predetermined range. The weights of each layer of the long short-term memory network (forget gate, input gate, memory unit and output gate) change during the training process. Each layer is the same at the beginning, and the subsequent training situation depends on the training situation. The process of training the long short-term memory network is a conventional training method of the existing neural network model, which will not be described here.

[0125] In actual application scenarios, the pre-construction of the third neural network model may include:

[0126] With x1, x2, …, x t represents the input data, x t is the input vector at time t, and the hidden unit h at the previous moment t-1 The parameters to be learned are mathematically calculated to obtain the hidden unit h at the current moment. t , and finally the training of the long short-term memory network is realized, and the actual values ​​of the parameters to be learned of the long short-term memory network are obtained.

[0127] The first neural network model and the second neural network model are both obtained by adding two Sigmoid gating units to the third neural network model, that is, an optimized long short-term memory network formed by adding two Sigmoid gating units to the long short-term memory network, including:

[0128] On the basis of the long short-term memory network architecture, the hidden unit at the previous moment is combined with a Sigmoid gating unit to obtain the optimized hidden unit at the previous moment, the input of the long short-term memory network is combined with another Sigmoid gating unit to obtain the optimized input of the long short-term memory network, and the optimized hidden unit at the previous moment and the optimized input of the long short-term memory network are respectively subjected to linear matrix transformation and combined to obtain the cell state; wherein, the Sigmoid gating unit combined with the hidden unit at the previous moment is connected to the input of the long short-term memory network, and the Sigmoid gating unit combined with the input of the long short-term memory network is connected to the optimized hidden unit at the previous moment.

[0129] The pre-construction of the first neural network model and the second neural network model includes:

[0130] Obtain historical substation-level data and city-level data to form a substation-level data training set and a city-level data training set respectively; wherein the substation-level data includes the substation load element data of each type of substation, the meteorological data of the area corresponding to the substation, and the actual load results of all types of substations; the city-level data includes the actual load results of all types of substations, the city-level load element data and the city-level meteorological data, and the actual load results of all cities;

[0131] Based on the substation-level data training set and the city-level data training set, the optimized long short-term memory network is trained until convergence, the parameters to be learned in the optimized long short-term memory network are updated, and the first neural network model and the second neural network model are obtained respectively.

[0132] The first neural network model and the second neural network model are both optimized long short-term memory networks, and both satisfy the following relationship:

[0133] u t =σ(W u x t +b u )

[0134] v t =σ(W v h t-1 +b v )

[0135] h′ t-1 =2h t-1 ⊙u t

[0136] x′ t =2x t ⊙v t

[0137] f t =σ(W hf h′ t-1 +W xf x′ t +b f )

[0138] i t =σ(W hi h′ t-1 +W xi x′ t +b i )

[0139]

[0140] o t =σ(W ho h′ t-1 +W xo x′t +b o )

[0141] h t =o t ⊙tanh(c t )

[0142] In the formula, u t and v t are the Sigmoid gated unit combined with the hidden unit of the previous moment and the Sigmoid gated unit combined with the input of the long short-term memory network, respectively. u and w v are the linear weight matrices for linear matrix transformation of the optimized hidden unit and the optimized input of the long short-term memory network at the previous moment, and b u and b v are the linear bias matrices when the linear matrix transformation is performed on the optimized hidden unit of the previous moment and the optimized input of the long short-term memory network, x t and x' t are the input and optimization input of the long short-term memory network, respectively, h t-1 and h' t-1 are the hidden units at the previous moment and the optimized hidden units, respectively, t 、i t , o t are the output values ​​of the forget gate, input gate, and output gate respectively. is the memory unit before updating, c t is the updated memory unit, c t-1 is the memory unit updated at the last moment, [W hf W xf ]、[W hi W xi ]、[W hg W xg ]、[W ho W xo ] are the network weights of the forget gate, input gate, memory unit before update and output gate, respectively. f , b i , b g , b o are the network biases of the forget gate, input gate, memory unit before update, and output gate, σ() is the Sigmoid logic activation function, ⊙ is the Hadamard product operator of the matrix, and h t is the hidden unit at the current moment, that is, the output result of the optimized long short-term memory network at time t. Figure 3 As shown, the dotted box is the added Sigmoid gating unit u t and v tThe third neural network model directly uses the input of the long short-term memory network and the hidden unit of the previous moment to calculate the hidden unit of the current moment, that is, the output of the long short-term memory network.

[0143] When the present invention performs load forecasting, the new energy output corresponding to the substation and the meteorological data of the area corresponding to the substation, and the new energy output in the city and the city-level meteorological data are respectively used as the first neural network model and the second neural network model after forming correlation coefficients;

[0144] The following relationships are satisfied between the renewable energy output corresponding to the substation and the meteorological data of the area corresponding to the substation, as well as between the renewable energy output in the city and the municipal meteorological data:

[0145]

[0146] Among them, r is the correlation coefficient between energy output and meteorological data, x i ,y i They are the new energy output value and meteorological data value within a limited time period. are the average values ​​of the new energy output value and the meteorological data value within the limited time period, and n is the number of new energy output values ​​or meteorological data values ​​within the limited time period. When calculating r, select the same number of x and y.

[0147] After the first neural network model, the second neural network model and the third neural network model are pre-built, the corresponding data and correlation coefficients can be input for load forecasting.

[0148] The present invention is based on provincial, municipal and substation load data, takes into account meteorological factors and socio-economic factors, analyzes load elements at each level, and adopts an improved clustering algorithm and a three-layer stacked improved long short-term memory neural network algorithm prediction model to complete the bottom-up prediction of multi-level loads, which can significantly improve the prediction accuracy of multi-level power loads and has application value for load prediction.

[0149] In the actual application scenario, the LSTM network is trained based on the load data, meteorological data and regional economic data of S Province, Y City and Y City 220kV substation from 2019 to June 2023, and load forecasting is performed to predict the average load in July 2023 to verify the effectiveness of the present invention.

[0150] First, loads of multiple substations in the first layer are clustered and divided into four categories with similar user electricity consumption behavior characteristics, such as Figure 4 :

[0151] Taking the summer of 2022 as an example, the correlation between photovoltaic output and weather in City Y is as follows: Figure 5As shown in Figure 1, meteorological data include temperature (℃), humidity (%), air pressure (hPa), ground wind speed (m / s), precipitation (mm / h), surface horizontal radiation (W / m2), etc. The Pearson correlation coefficient between distributed photovoltaic output and various meteorological physical quantities is calculated. Table 1 shows that the output of distributed photovoltaic has a strong positive correlation with temperature and surface horizontal radiation, a strong negative correlation with humidity, and a weak correlation with air pressure.

[0152] Table 1 Pearson correlation coefficients between summer photovoltaic output and meteorological physical quantities

[0153]

[0154]

[0155] Based on the clustering results of substations, the first-layer improved LSTM neural network algorithm (first neural network model) is trained, and the input data is the load data of Y city substations from June 2019 to 2023, new energy output data, substation capacity-load ratio and load rate, meteorological data, etc. In the load forecast, the capacity-load ratio constraint of the 220kV substation is based on the actual capacity-load ratio and is set in the range of 1.6 to 1.9; the load rate constraint is not higher than 75%, and the natural growth of the load is considered.

[0156] The second-layer improved LSTM neural network algorithm model (the second neural network model) takes a certain day in June as an example, and the input data structure of the full network load of City Y is shown in Table 2; the third-layer LSTM neural network algorithm model (the third neural network model), some industrial data of Province S are shown in Table 3.

[0157] Table 2 Y city load data structure

[0158]

[0159]

[0160] Table 3 Data of some industries in S Province

[0161]

[0162] The trained three-layer load forecasting model is used to predict the average load in July 2023 and the maximum load value in 2023, and compared with the prediction results of the actual load and the control test algorithm model. Figure 6 , Figure 7 They are the average loads of Y City and S Province in July respectively. Table 4 is the error comparison of the three load forecasting models.

[0163] The present invention considers meteorological factors and socio-economic data and adopts an improved LSTM neural network algorithm, which is recorded as FD-LSTM. The control experiment selects ① not considering meteorological factors and socio-economic data, adopting an improved LSTM neural network algorithm, which is recorded as DYN-LSTM; ② considering meteorological factors and socio-economic data, adopting a conventional LSTM neural network algorithm, which is recorded as FAC-LSTM.

[0164] Table 4 Average load forecast error in Y City and S Province in July

[0165]

[0166]

[0167] It can be seen that the prediction accuracy of the prediction model (FD-LSTM) that adopts the improved LSTM neural network algorithm and considers meteorological factors and socio-economic data is higher than that of the prediction model (DYN-LSTM) that does not consider meteorological factors and socio-economic data; and the accuracy of the prediction model (FD-LSTM) that also considers meteorological factors and socio-economic data is higher than that of the prediction model (FAC-LSTM) that adopts the conventional LSTM neural network algorithm. Therefore, the prediction results of the multi-level power load prediction method based on improved K-Means and LSTM of the present invention have smaller errors with the actual load and higher prediction accuracy.

[0168] like Figure 8 As shown, the present invention also provides a substation multi-level power load forecasting system, which adopts the above-mentioned substation multi-level power load forecasting method, and the system includes:

[0169] A data acquisition module, which is used to acquire city-level meteorological data and provincial-level economic data, and acquire substation load element data of each category of substation based on the clustering results of pre-optimized clustering of all substations;

[0170] A substation load prediction module is used to perform load prediction based on a pre-built first neural network model by inputting substation load element data of each type of substation and meteorological data of the area corresponding to the substation, and providing load prediction results for all types of substations;

[0171] A municipal load forecasting module is used to perform load forecasting based on a pre-built second neural network model by inputting load forecasting results of all types of substations, municipal load factor data and municipal meteorological data, and providing load forecasting results of all municipalities;

[0172] The provincial load forecasting module is used to perform load forecasting based on the pre-built third neural network model, input all city-level load forecasting results, provincial load factor data and provincial economic data, and give provincial load forecasting results.

[0173] Furthermore, the substation multi-level power load forecasting system also includes a data processing module, which includes:

[0174] The time series relationship of multi-level power loads is established, and the first neural network model, the second neural network model and the third neural network are modified.

[0175] Furthermore, the data processing module also includes:

[0176] Based on the time series relationship between the load data of the substation and the load data of all districts and counties in the city, the actual load results at the city level are determined through the actual load results of the substation;

[0177] Based on the time series relationship between the load history data of all districts and counties in the city and the load history data of all prefecture-level cities in the province, the actual load results at the provincial level are determined through the actual load results at the municipal level;

[0178] The actual load results of the substation, the actual load results of the municipal level and the actual load results of the provincial level are applied to the training of the first neural network model, the second neural network model and the third neural network.

[0179] Furthermore, the data acquisition module includes:

[0180] Resample the data set containing the historical load curves of all substations in the province to obtain multiple new data sets;

[0181] Cluster each new data set and give the cluster center corresponding to each new data set;

[0182] All cluster centers are combined to form a central data set, and hierarchical clustering is performed to obtain multiple clusters;

[0183] Traverse the original data set containing the historical load curves of all substations in the province and the central data set, and match each historical load curve with the cluster center closest to it from the central data set;

[0184] Based on the matching results, the historical load curves in the original data set are divided into clusters corresponding to the matching cluster centers;

[0185] Based on the clusters divided by historical load curves and their corresponding relationships with substations, the results of substation division into different categories are obtained.

[0186] Furthermore, the substation multi-level power load forecasting system also includes a model building module, which includes:

[0187] Obtain historical provincial data to form a provincial data training set; provincial data includes all city-level actual load results, provincial load factor data, provincial economic data and provincial actual load results;

[0188] Based on the provincial data training set, the long short-term memory network is trained until convergence, and the parameters to be learned in the long short-term memory network are updated to obtain the third neural network model.

[0189] Furthermore, the model building module includes:

[0190] Obtain historical substation-level data and city-level data to form a substation-level data training set and a city-level data training set respectively; wherein the substation-level data includes the substation load element data of each type of substation, the meteorological data of the area corresponding to the substation, and the actual load results of all types of substations; the city-level data includes the actual load results of all types of substations, the city-level load element data and the city-level meteorological data, and the actual load results of all cities;

[0191] Based on the substation-level data training set and the city-level data training set, the optimized long short-term memory network is trained until convergence, the parameters to be learned in the optimized long short-term memory network are updated, and the first neural network model and the second neural network model are obtained respectively.

[0192] Furthermore, the model building module also includes:

[0193] On the basis of the long short-term memory network architecture, the hidden unit at the previous moment is combined with a Sigmoid gating unit to obtain the optimized hidden unit at the previous moment, the input of the long short-term memory network is combined with another Sigmoid gating unit to obtain the optimized input of the long short-term memory network, and the optimized hidden unit at the previous moment and the optimized input of the long short-term memory network are respectively subjected to linear matrix transformation and combined to obtain the cell state; wherein, the Sigmoid gating unit combined with the hidden unit at the previous moment is connected to the input of the long short-term memory network, and the Sigmoid gating unit combined with the input of the long short-term memory network is connected to the optimized hidden unit at the previous moment.

[0194] Although preferred embodiments of the present invention have been described, additional changes and modifications may be made to these embodiments by those skilled in the art once the basic inventive concepts are known. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A multi-level power load forecasting method for a substation, characterized in that: include: Based on the clustering results of pre-optimized clustering of all substations, substation load element data of each category of substations are obtained, and first-level regional economic data and second-level regional meteorological data are obtained; wherein the first-level region includes multiple second-level regions, and the second-level region includes multiple substations; Based on the pre-built first neural network model, the substation load factor data of each type of substation and the meteorological data of the area corresponding to the substation are input to perform load forecasting, and the load forecast results of all types of substations are given; Based on the pre-built second neural network model, the load forecast results of all types of substations, the second-level regional load factor data and the second-level regional meteorological data are input to perform load forecasting, and the load forecast results of all second-level regions are given; Based on the pre-built third neural network model, the load forecast results of all second-level regions, the load factor data of the first-level regions and the economic data of the first-level regions are input for load forecasting, and the load forecast results of the first-level regions are given.

2. The multi-level power load forecasting method for substations according to claim 1, characterized in that: The substation load element data includes the substation load historical data, the substation's capacity load ratio and load rate, the substation line structure, and the corresponding new energy output of the substation. The second-level regional load element data includes the second-level regional load historical data, the second-level regional industrial structure, and the second-level regional new energy output. The first-level regional load element data includes the first-level regional load historical data, the first-level regional industrial structure, and the first-level regional new energy output.

3. The multi-level power load forecasting method for substations according to claim 2, characterized in that: When the first neural network model, the second neural network model and the third neural network model are pre-built, it also includes: The time series relationship of multi-level power loads is established, and the first neural network model, the second neural network model and the third neural network are modified.

4. The multi-level power load forecasting method for substations according to claim 3, characterized in that: The time series relationship of multi-level power loads includes the time series relationship between the substation load data and the load data of the second-level region, and the time series relationship between the load history data of the second-level region and the load history data of the first-level region, all of which satisfy the following relationship: In the formula, y is the upper load data, x1, x2, ..., x n is the n lower-level load data corresponding to the upper level, n is the number of lower-level loads, a i is the correction coefficient corresponding to the ith lower-level load data, and c is a constant; The first neural network model, the second neural network model and the third neural network are modified, including: Based on the time series relationship between the load data of the substation and the historical load data of the second-level area, the actual load result of the second-level area is determined by the actual load result of the substation; Based on the time series relationship between the load history data of the second-level region and the load history data of the first-level region, the actual load result of the first-level region is determined by the actual load result of the second-level region; The actual load results of the substation, the actual load results of the second-level area, and the actual load results of the first-level area are applied to the training of the first neural network model, the second neural network model, and the third neural network.

5. The multi-level power load forecasting method for substations according to claim 1, characterized in that: All substations are optimized and clustered in advance, including: Resampling the data set containing the historical load curves of all substations in the first-level area to obtain multiple new data sets; Cluster each new data set and give the cluster center corresponding to each new data set; All cluster centers are combined to form a central data set, and hierarchical clustering is performed to obtain multiple clusters; Traverse the original data set containing the historical load curves of all substations in the first-level area and the central data set, and match each historical load curve with the cluster center closest to it from the central data set; Based on the matching results, the historical load curves in the original data set are divided into clusters corresponding to the matching cluster centers; Based on the clusters divided by historical load curves and their corresponding relationships with substations, the results of substation division into different categories are obtained.

6. The multi-level power load forecasting method for substations according to claim 5, characterized in that: Pre-built third-party neural network models, including: Obtain historical first-level regional data to form a first-level regional data training set; the first-level regional data includes actual load results of all second-level regions, first-level regional load element data, first-level regional economic data and actual load results of the first-level regions; The long short-term memory network is trained based on the first-level regional data training set until convergence, and the parameters to be learned in the long short-term memory network are updated to obtain the third neural network model.

7. The multi-level power load forecasting method for substations according to claim 6, characterized in that: The architectures of the first neural network model and the second neural network model are both optimized long short-term memory networks formed by adding two Sigmoid gating units to the long short-term memory network; The pre-construction of the first neural network model and the second neural network model includes: Obtain historical substation-level data and second-level regional data to form a substation-level data training set and a second-level regional data training set respectively; wherein the substation-level data includes substation load element data of each category of substation, meteorological data of the area corresponding to the substation, and actual load results of all categories of substations, and the second-level regional data includes actual load results of all categories of substations, second-level regional load element data, second-level regional meteorological data, and actual load results of all second-level regions; Based on the substation-level data training set and the second-level regional data training set, the optimized long short-term memory network is trained until convergence, the parameters to be learned in the optimized long short-term memory network are updated, and the first neural network model and the second neural network model are obtained respectively.

8. The multi-level power load forecasting method for substations according to claim 7, characterized in that: Add two Sigmoid gate units to the long short-term memory network, including: On the basis of the long short-term memory network architecture, the hidden unit at the previous moment is combined with a Sigmoid gating unit to obtain the optimized hidden unit at the previous moment, the input of the long short-term memory network is combined with another Sigmoid gating unit to obtain the optimized input of the long short-term memory network, and the optimized hidden unit at the previous moment and the optimized input of the long short-term memory network are respectively subjected to linear matrix transformation and combined to obtain the cell state; wherein, the Sigmoid gating unit combined with the hidden unit at the previous moment is connected to the input of the long short-term memory network, and the Sigmoid gating unit combined with the input of the long short-term memory network is connected to the optimized hidden unit at the previous moment.

9. The multi-level power load forecasting method for substations according to claim 8, characterized in that: Optimize the long short-term memory network to satisfy the following relationship: you t =σ(W u x t +b u ) v t =σ(W v h t-1 +b v ) <h2 style=";text-align:left;direction:ltr">h′<h2 style=";text-align:left;direction:ltr"> t-1 <h2 style=";text-align:left;direction:ltr"> <2h<h2 style=";text-align:left;direction:ltr"> t-1 <h2 style=";text-align:left;direction:ltr"> ⊙u<h2 style=";text-align:left;direction:ltr"> t x′ t =2x t ⊙v t f t =σ(W hf h′ t-1 +W xf x′ t +b f ) i t =σ(W hi h′ t-1 +W xi x′ t +b i ) o t =σ(W ho h′ t-1 +W xo x′ t +b o ) h t =o t ⊙tanh(c t ) In the formula, u t and v t are the Sigmoid gated unit combined with the hidden unit of the previous moment and the Sigmoid gated unit combined with the input of the long short-term memory network, respectively. u and w v are the linear weight matrices for linear matrix transformation of the optimized hidden unit and the optimized input of the long short-term memory network at the previous moment, and b u and b v are the linear bias matrices when the linear matrix transformation is performed on the optimized hidden unit of the previous moment and the optimized input of the long short-term memory network, x t and x' t are the input and optimization input of the long short-term memory network, respectively, h t-1 and h' t-1 are the hidden units at the previous moment and the optimized hidden units, respectively, t 、i t , o t are the output values ​​of the forget gate, input gate, and output gate respectively. is the memory unit before updating, c t is the updated memory unit, c t-1 is the memory unit updated at the last moment, [W hf W xf ]、[W hi W xi ]、[W hg W xg ]、[W ho W xo ] are the network weights of the forget gate, input gate, memory unit before update and output gate, respectively. f , b i , b g , b o are the network biases of the forget gate, input gate, memory unit before update, and output gate, σ() is the Sigmoid logic activation function, ⊙ is the Hadamard product operator of the matrix, and h t is the hidden unit at the current moment.

10. A multi-level power load forecasting system for a substation, characterized in that: The multi-level power load forecasting method for a substation as claimed in any one of claims 1 to 9 is adopted, and the system comprises: A data acquisition module, which is used to acquire the second-level regional meteorological data and the first-level regional economic data, and acquire the substation load element data of each category of substation based on the clustering result of pre-optimized clustering of all substations; A substation load prediction module is used to perform load prediction based on a pre-built first neural network model by inputting substation load element data of each type of substation and meteorological data of the area corresponding to the substation, and providing load prediction results for all types of substations; The second-level regional load forecasting module is used to perform load forecasting based on the pre-built second neural network model, input the load forecasting results of all types of substations, the second-level regional load factor data and the second-level regional meteorological data, and provide the load forecasting results of all second-level regions; The first-level regional load forecasting module is used to perform load forecasting based on the pre-built third neural network model, input the load forecasting results of all second-level regions, the first-level regional load factor data and the first-level regional economic data, and give the load forecasting results of the first-level regions.

Images