Power grid load prediction method and device based on multiple models

Abstract

The invention discloses a power grid load prediction method and device based on multiple models. The method comprises the steps of obtaining historical load data of multiple substations, forming initial historical load data, performing clustering analysis on the initial historical load data, and giving category division of the initial historical load data; performing load prediction on each transformer substation through an improved LSTM model and the initial historical load data of each category, and performing load prediction on the power network of the first region in combination with the historical load data of the first region and the industrial data and meteorological data of the first region; and based on the load prediction of the first regional power network, performing load prediction on the second regional power network through the constructed neural network combination model in combination with the second regional historical load data. According to the method, the power grid load prediction for the multi-stage power network is realized through multiple models from bottom to top, each model considers the characteristics of each stage of power network, and the prediction precision of each stage of power grid load is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power grid load, and in particular relates to a power grid load prediction method and device based on multiple models. Background Art

[0002] Grid load forecasting is a very important part of power system planning and operation. It is of great significance to ensure the economy, reliability and safety of the power system. Grid load forecasting needs to combine economic, meteorological and grid structure information through forecasting models.

[0003] For example, patent CN118944084A provides a method and device for predicting power load, and constructs a power load prediction model architecture, wherein the power load prediction model architecture is composed of multiple linear prediction models, and each prediction model corresponds to a preset weight; based on multiple optimization algorithms, the preset weights of each prediction model in the power load prediction model architecture are optimized to obtain an optimization result data set; based on the prediction accuracy corresponding to each optimized power load prediction model in the optimization result data set and the pre-constructed weight setting rules, the final power load prediction model is determined; based on the final power load prediction model, the power load of the target area and the target period is predicted. The power load prediction model provided by this solution can improve the accuracy of power load.

[0004] However, with the construction of new power systems, the high proportion of new energy access and its inherent random volatility make the load changes more complicated. The factors caused by the proportion of new energy access are random and changeable, and the nonlinear relationship between power load and influencing factors is complex. It is difficult to express the relationship between the two with an accurate relationship, so a new and more accurate load forecasting method is needed.

[0005] Some studies have used a combination model to achieve power load forecasting under multiple factors. For example, patent CN117522443A provides a method and system for predicting the day-ahead electricity price in the power market based on CNN-LSTM-Attention, including dividing the standardized historical electricity price information data into a training set and a test set; inputting the training set into the CNN-LSTM model, performing parameter optimization learning, and obtaining an optimized CNN-LSTM model; automatically performing weighted average calculation on the output data of the optimized CNN-LSTM model through the attention mechanism to obtain a CNN-LSTM-Attention model; and predicting the day-ahead electricity price on the forecast day through the CNN-LSTM-Attention model. This solution can ensure the accuracy of the electricity price forecast results and help power companies to reasonably formulate electricity trading strategies based on the electricity price forecast results.

[0006] Currently, there is no unified systematic research on power networks at all levels to achieve simultaneous prediction of grid loads at all levels of power networks.

[0007] Therefore, how to simultaneously predict the grid loads of power networks at all levels to achieve high-precision prediction of grid loads at all levels is an urgent problem to be solved by technical personnel in this field. Summary of the invention

[0008] In view of the defects existing in the above-mentioned prior art, the present invention provides a power grid load forecasting method and device based on multiple models. The power grid load forecasting method specifically includes the following steps: obtaining historical load data of multiple substations to form initial historical load data, and performing cluster analysis on the initial historical load data to provide category division of the initial historical load data; through an improved LSTM model, based on the initial historical load data of each category, performing load forecasting on each substation, and combining the historical load data of the first area and the industrial data and meteorological data of the first area, performing load forecasting on the power network of the first area; based on the load forecast of the power network of the first area, through the constructed neural network combination model, combined with the historical load data of the second area, performing load forecasting on the power network of the second area; wherein, the power network of the first area includes multiple substations, and the power network of the second area includes multiple power networks of the first area.

[0009] The present invention realizes grid load prediction for multi-level power networks through multiple models in a bottom-up manner. Each model takes into account the characteristics of each level of the power network, which significantly improves the prediction accuracy of the grid load at each level.

[0010] In a first aspect, the present invention provides a method for predicting power grid load based on multiple models, which specifically comprises the following steps:

[0011] Obtain historical load data of multiple substations to form initial historical load data, perform cluster analysis on the initial historical load data, and provide category division of the initial historical load data;

[0012] Through the improved LSTM model, based on the initial historical load data of each category, the load of each substation is predicted, and the load of the power network in the first area is predicted by combining the historical load data of the first area and the industrial data and meteorological data of the first area;

[0013] Based on the load forecast of the power network in the first area, the load forecast of the power network in the second area is performed by combining the historical load data of the second area with the constructed neural network combination model;

[0014] The first regional power network includes a plurality of substations, and the second regional power network includes a plurality of first regional power networks.

[0015] Further, the first regional power network is a municipal power network, and the second regional power network is a provincial power network;

[0016] Furthermore, historical load data of multiple substations are obtained to form initial historical load data, and cluster analysis is performed on the initial historical load data to provide a type classification of the initial historical load data, specifically including:

[0017] Preprocess the historical load data of multiple substations to form historical load curves of multiple time frames;

[0018] Resample each historical load curve to form a corresponding load sample;

[0019] By clustering multiple load samples, multiple groups of clusters are given;

[0020] Based on the association analysis of each group of clusters, multiple clusters are formed;

[0021] By comparing the historical load data of each substation with each cluster center, the historical load data of each substation is classified into the corresponding cluster to complete the type classification of the initial historical load data.

[0022] Furthermore, each load curve is resampled to form a corresponding load sample, including:

[0023] Define the number of samples for resampling, the number of resampling times, and the statistical characteristics of resampling, perform statistical characteristic analysis on each resampled sample, fuse each statistical characteristic, perform deviation correction, set the confidence interval, and form the corresponding load sample;

[0024] Based on the association analysis of each group of clusters, multiple clusters are formed, including:

[0025] Calculate the centroid distance between each group of clusters, which is specifically expressed as:

[0026]

[0027] Among them, μ i ,μ j Cluster C i With cluster C j The center of mass, d(μ i ,μ j ) is cluster C i With cluster C j The distance between the centroids, μ ik ,μ jk They are μ in k dimension respectively. i and μ j The coordinates of the cluster, d is the dimension endpoint, n is the total number of clusters, x ik ,xjk Cluster C i With cluster C j Elements of

[0028] Based on the comparison with the threshold distance, clusters are merged to form cluster groups;

[0029] The cluster group is removed and updated into a new group of clusters, and the centroid distance calculation between each group of clusters is repeated until the association analysis of each group of clusters is completed, and the remaining groups of clusters correspond to each cluster.

[0030] Furthermore, through the improved LSTM model, based on the initial historical load data of each category, the load forecast of each substation is carried out, and the load forecast of the power network in the first area is carried out in combination with the historical load data of the first area and the industrial data and meteorological data of the first area, specifically including:

[0031] By matching each substation with the initial historical load data of each category, combined with the substation attribute data and meteorological data, and using the improved LSTM model, the load forecast data of each substation is given;

[0032] The load forecast data of the substation, the historical load data of the first region, the industrial data of the first region, and the meteorological data are integrated, and the load forecast of the power network in the first region is carried out through the improved LSTM model.

[0033] Furthermore, the improved LSTM model includes adding a sigmoid neural network layer group in the LSTM model architecture, the sigmoid neural network layer group including a first sigmoid neural network layer, a second sigmoid neural network layer and a fusion layer;

[0034] The first sigmoid neural network layer is used to control the input at the current moment and realize the Schur product calculation of the output value and the output hidden state at the previous moment to form the optimized output hidden state at the previous moment;

[0035] The second sigmoid neural network layer is used to control the optimized hidden state of the previous moment output and calculate the Schur product of the output value and the current moment input to form the optimized current moment input;

[0036] The fusion layer is used to combine the weights and bias parameters to fuse the optimized hidden state of the previous moment output and the optimized input of the current moment, and then transmit them to the forget gate, input gate and output gate in the LSTM model architecture respectively.

[0037] Furthermore, the first sigmoid neural network layer is used to control the input at the current moment, and realizes the Schur product calculation of the output value and the output hidden state at the previous moment to form the optimized output hidden state at the previous moment, which is specifically expressed as:

[0038] u t =σ(W u x t +b u )

[0039] h' t-1 =h t-1 *u t

[0040] Among them, u t is the output value at time t of the first sigmoid neural network layer, x t is the input at time t, W u is the weight of the first sigmoid neural network layer, b u is the bias parameter of the first sigmoid neural network layer, σ() is the Sigmoid activation function, and h t-1 Output hidden state at time t-1, h' t-1 is the optimized output hidden state at time t-1, * is the Schur product operation;

[0041] Furthermore, the second sigmoid neural network layer is used to control the optimized output hidden state of the previous moment, and realizes the calculation of the Schur product of the output value and the current moment input to form the optimized current moment input, which is specifically expressed as:

[0042] v t =σ(W v h t-1 '+bv)

[0043] x' t =x t *v t

[0044] Among them, v t is the output value at time t of the second sigmoid neural network layer, h t-1 Output hidden state at time t-1, W v is the weight of the second sigmoid neural network layer, b v is the bias parameter of the second sigmoid neural network layer, σ() is the Sigmoid activation function, x' t is the optimized input at time t, x t Input at time t;

[0045] Furthermore, the fusion layer is used to fuse the optimized hidden state of the previous moment output and the optimized input of the current moment by combining the weight and bias parameters, which is specifically expressed as:

[0046] s t =(W 1 h' t-1 +b l1 )+(W 2 x' t +b l2 )

[0047] Among them, s t is the fusion output, W 1 , W 2 are the weights of the first sigmoid neural network layer and the second sigmoid neural network layer in the fusion layer, b l1 、b l2 are the bias parameters corresponding to the first sigmoid neural network layer and the second sigmoid neural network layer in the fusion layer.

[0048] Furthermore, the attribute data of the substation includes the capacity-load ratio, load rate and new energy output data of the substation;

[0049] By matching each substation with the initial historical load data of each category, combined with the attribute data and meteorological data of the substation, the load forecast data of each substation is given through the improved LSTM model, including:

[0050] The time series is determined by the initial historical load data of each substation, and the capacity-load ratio, load rate, new energy processing data, and meteorological data of the substation are processed according to the time series to form multi-dimensional feature data;

[0051] Sort out the input quantity, time series length, and feature dimension values ​​to form a three-dimensional input tensor;

[0052] The three-dimensional input tensor is fed into the improved LSTM model to give the load forecast data for each substation.

[0053] Furthermore, the load forecast data of the substation, the historical load data of the first region, the industrial data of the first region, and the meteorological data are integrated to perform load forecasting on the power network of the first region through the improved LSTM model, including:

[0054] Determine a time series using the historical load data of the first region, and process the load forecast data of the corresponding power station, the industrial data of the first region, and the meteorological data according to the time series to form multi-dimensional feature data;

[0055] Sort out the input quantity, time series length, and feature dimension values ​​to form a three-dimensional input tensor;

[0056] The three-dimensional input tensor is fed into the improved LSTM model to give the load forecast data of the power network in the first region.

[0057] Furthermore, the construction of the neural network combination model specifically includes:

[0058] Multiple groups of LSTM models are used in parallel to build a corresponding multi-channel structure, where each group of LSTM models includes multiple LSTM model units;

[0059] The multi-channel structure is connected to the transposition layer, which is connected to the CNN model. The convolutional layer in the CNN model performs feature fusion and gives the prediction result.

[0060] Furthermore, based on the load forecast of the power network in the first area, the load forecast of the power network in the second area is performed by combining the constructed neural network combination model with the historical load data of the second area, specifically including:

[0061] Determine the time series with the historical load data of the second region, and process the load forecast data of the power network corresponding to the first region, the industrial data of the second region, and the meteorological data according to the time series to form multi-dimensional feature data, wherein the dimension of the feature data is the same as the number of groups of the LSTM model in the neural network combination model;

[0062] Input each dimension feature data into the LSTM model of the corresponding group respectively, and output the initial result sequence;

[0063] The initial structure sequence of each group of outputs is transposed and then input into the CNN model. After the feature fusion of the convolution layer, the load forecast data of the power network in the second area is given.

[0064] Among them, the feature fusion of the convolutional layer is specifically expressed as:

[0065]

[0066] Among them, S (1,n) is the output after feature fusion of the convolutional layer of the CNN model, f() is the activation function, b is the bias vector of the convolutional layer of the CNN model, M is the number of LSTM model groups, N is the number of LSTM model units in each group of LSTM models, and H s,p is the initial result output by the pth LSTM model unit in the sth group of LSTM models, ω s,p is the corresponding convolution kernel weight matrix.

[0067] In a second aspect, the present invention further provides a power grid load forecasting device based on multiple models, specifically comprising:

[0068] The acquisition module is used to obtain historical load data of multiple substations, form initial historical load data, obtain historical load data of the first area, industrial data and meteorological data of the first area, and historical load data of the second area;

[0069] An analysis module is used to perform cluster analysis on the initial historical load data and provide a classification of the initial historical load data;

[0070] The prediction module is used to perform load prediction for each substation based on the initial historical load data of each category through the improved LSTM model, and to perform load prediction for the power network in the first area in combination with the historical load data of the first area and the industrial data and meteorological data of the first area; based on the load prediction of the power network in the first area, the load prediction for the power network in the second area is performed through the constructed neural network combination model in combination with the historical load data of the second area.

[0071] Further, the first regional power network is a municipal power network, and the second regional power network is a provincial power network;

[0072] Furthermore, historical load data of multiple substations are obtained to form initial historical load data, and cluster analysis is performed on the initial historical load data to provide a type classification of the initial historical load data, specifically including:

[0073] Preprocess the historical load data of multiple substations to form historical load curves of multiple time frames;

[0074] Resample each historical load curve to form a corresponding load sample;

[0075] By clustering multiple load samples, multiple groups of clusters are given;

[0076] Based on the association analysis of each group of clusters, multiple clusters are formed;

[0077] By comparing the historical load data of each substation with each cluster center, the historical load data of each substation is classified into the corresponding cluster to complete the type classification of the initial historical load data.

[0078] Furthermore, each load curve is resampled to form a corresponding load sample, including:

[0079] Define the number of samples for resampling, the number of resampling times, and the statistical characteristics of resampling, perform statistical characteristic analysis on each resampled sample, fuse each statistical characteristic, perform deviation correction, set the confidence interval, and form the corresponding load sample;

[0080] Based on the association analysis of each group of clusters, multiple clusters are formed, including:

[0081] Calculate the centroid distance between each group of clusters, which is specifically expressed as:

[0082]

[0083] Among them, μ i ,μ j Cluster C i With cluster C j The center of mass, d(μ i ,μ j ) is cluster C i With cluster C j The distance between the centroids, μ ik ,μ jk They are μ in k dimension respectively. i and μ j The coordinates of the cluster, d is the dimension endpoint, n is the total number of clusters, x ik ,x jk Cluster C i With cluster C j Elements of

[0084] Based on the comparison with the threshold distance, clusters are merged to form cluster groups;

[0085] The cluster group is removed and updated into a new group of clusters, and the centroid distance calculation between each group of clusters is repeated until the association analysis of each group of clusters is completed, and the remaining groups of clusters correspond to each cluster.

[0086] Furthermore, through the improved LSTM model, based on the initial historical load data of each category, the load forecast of each substation is carried out, and the load forecast of the power network in the first area is carried out in combination with the historical load data of the first area and the industrial data and meteorological data of the first area, specifically including:

[0087] By matching each substation with the initial historical load data of each category, combined with the substation attribute data and meteorological data, and using the improved LSTM model, the load forecast data of each substation is given;

[0088] The load forecast data of the substation, the historical load data of the first region, the industrial data of the first region, and the meteorological data are integrated, and the load forecast of the power network in the first region is carried out through the improved LSTM model.

[0089] Furthermore, the improved LSTM model includes adding a sigmoid neural network layer group in the LSTM model architecture, the sigmoid neural network layer group including a first sigmoid neural network layer, a second sigmoid neural network layer and a fusion layer;

[0090] The first sigmoid neural network layer is used to control the input at the current moment and realize the Schur product calculation of the output value and the output hidden state at the previous moment to form the optimized output hidden state at the previous moment;

[0091] The second sigmoid neural network layer is used to control the optimized hidden state of the previous moment output and calculate the Schur product of the output value and the current moment input to form the optimized current moment input;

[0092] The fusion layer is used to combine the weights and bias parameters to fuse the optimized hidden state of the previous moment output and the optimized input of the current moment, and then transmit them to the forget gate, input gate and output gate in the LSTM model architecture respectively.

[0093] Furthermore, the first sigmoid neural network layer is used to control the input at the current moment, and realizes the Schur product calculation of the output value and the output hidden state at the previous moment to form the optimized output hidden state at the previous moment, which is specifically expressed as:

[0094] u t =σ(W u x t +b u )

[0095] h' t-1 =h t-1 *u t

[0096] Among them, u t is the output value at time t of the first sigmoid neural network layer, x t is the input at time t, W u is the weight of the first sigmoid neural network layer, b u is the bias parameter of the first sigmoid neural network layer, σ() is the Sigmoid activation function, and h t-1 Output hidden state at time t-1, h' t-1 is the optimized output hidden state at time t-1, * is the Schur product operation;

[0097] Furthermore, the second sigmoid neural network layer is used to control the optimized output hidden state of the previous moment, and realizes the calculation of the Schur product of the output value and the current moment input to form the optimized current moment input, which is specifically expressed as:

[0098] v t =σ(W v h t-1 '+b v )

[0099] x't =x t *v t

[0100] Among them, v t is the output value at time t of the second sigmoid neural network layer, h t-1 Output hidden state at time t-1, W v is the weight of the second sigmoid neural network layer, b v is the bias parameter of the second sigmoid neural network layer, σ() is the Sigmoid activation function, x' t is the optimized input at time t, x t Input at time t;

[0101] Furthermore, the fusion layer is used to fuse the optimized hidden state of the previous moment output and the optimized input of the current moment by combining the weight and bias parameters, which is specifically expressed as:

[0102] s t =(W 1 h' t-1 +b l1 )+(W 2 x' t +b l2 )

[0103] Among them, s t is the fusion output, W 1 , W 2 are the weights of the first sigmoid neural network layer and the second sigmoid neural network layer in the fusion layer, b l1 、b l2 are the bias parameters corresponding to the first sigmoid neural network layer and the second sigmoid neural network layer in the fusion layer.

[0104] Furthermore, the attribute data of the substation includes the capacity-load ratio, load rate and new energy output data of the substation;

[0105] By matching each substation with the initial historical load data of each category, combined with the attribute data and meteorological data of the substation, the load forecast data of each substation is given through the improved LSTM model, including:

[0106] The time series is determined by the initial historical load data of each substation, and the capacity-load ratio, load rate, new energy processing data, and meteorological data of the substation are processed according to the time series to form multi-dimensional feature data;

[0107] Sort out the input quantity, time series length, and feature dimension values ​​to form a three-dimensional input tensor;

[0108] The three-dimensional input tensor is fed into the improved LSTM model to give the load forecast data for each substation.

[0109] Furthermore, the load forecast data of the substation, the historical load data of the first region, the industrial data of the first region, and the meteorological data are integrated to perform load forecasting on the power network of the first region through the improved LSTM model, including:

[0110] Determine a time series using the historical load data of the first region, and process the load forecast data of the corresponding power station, the industrial data of the first region, and the meteorological data according to the time series to form multi-dimensional feature data;

[0111] Sort out the input quantity, time series length, and feature dimension values ​​to form a three-dimensional input tensor;

[0112] The three-dimensional input tensor is fed into the improved LSTM model to give the load forecast data of the power network in the first region.

[0113] Furthermore, the construction of the neural network combination model specifically includes:

[0114] Multiple groups of LSTM models are used in parallel to build a corresponding multi-channel structure, where each group of LSTM models includes multiple LSTM model units;

[0115] The multi-channel structure is connected to the transposition layer, which is connected to the CNN model. The convolutional layer in the CNN model performs feature fusion and gives the prediction result.

[0116] Furthermore, based on the load forecast of the power network in the first area, the load forecast of the power network in the second area is performed by combining the constructed neural network combination model with the historical load data of the second area, specifically including:

[0117] Determine a time series using the historical load data of the second region, and process the load forecast data corresponding to the power network of the first region, the industrial data of the second region, and the meteorological data according to the time series to form multi-dimensional feature data, wherein the dimension of the feature data is the same as the number of groups of the LSTM model in the neural network combination model;

[0118] Input each dimension feature data into the LSTM model of the corresponding group respectively, and output the initial result sequence;

[0119] The initial structure sequence of each group of outputs is transposed and then input into the CNN model. After the feature fusion of the convolution layer, the load forecast data of the power network in the second area is given.

[0120] Among them, the feature fusion of the convolutional layer is specifically expressed as:

[0121]

[0122] Among them, S (1,n) is the output after feature fusion of the convolutional layer of the CNN model, f() is the activation function, b is the bias vector of the convolutional layer of the CNN model, M is the number of LSTM model groups, N is the number of LSTM model units in each group of LSTM models, and H s,p is the initial result output by the pth LSTM model unit in the sth group of LSTM models, ω s,p is the corresponding convolution kernel weight matrix.

[0123] The present invention provides a multi-model-based power grid load forecasting method and device, which at least have the following beneficial effects:

[0124] (1) The present invention implements grid load prediction for multi-level power networks through multiple models in a bottom-up manner. Each model takes into account the characteristics of each level of the power network, which significantly improves the prediction accuracy of the grid load at each level.

[0125] (2) Multiple groups of clusters are determined based on load samples, and then hierarchical association analysis is performed on the multiple groups of clusters. The hierarchical association analysis considers the centroid links of each group of clusters and determines the similarity between them by calculating the distance between the centroids of two clusters to form cluster groups. Finally, clusters with consistent categories are obtained, which are used as characteristic factors for load prediction of each substation.

[0126] (3) By improving the typical LSTM model, adding the first sigmoid neural network layer, the second sigmoid neural network layer and the fusion layer, the dynamic prediction capability is enhanced, and the accuracy of power grid load prediction at the substation level and the first regional power network (municipal level) is improved.

[0127] (4) Taking into account the complexity of characteristic factors in the grid load prediction of the second regional power network, the number of LSTM model groups of corresponding dimensions is set, and a corresponding number of LSTM model units are set in each group. The fusion of all characteristic factors is realized through the LSTM model group, and the intrinsic relationship between various characteristic factors and grid load is further explored from the output end, thereby improving the prediction effect of the grid load of the second regional power network. BRIEF DESCRIPTION OF THE DRAWINGS

[0128] Figure 1 A schematic diagram of a flow chart of a multi-model-based power grid load forecasting method provided by the present invention;

[0129] Figure 2 A schematic diagram of a flow chart for classifying the types of initial historical load data provided by a certain embodiment of the present invention;

[0130] Figure 3A schematic diagram of a flow chart of load forecasting for each substation and a first regional power network according to an embodiment of the present invention;

[0131] Figure 4 A network architecture diagram of an improved LSTM model of a certain embodiment provided by the present invention;

[0132] Figure 5 A network architecture diagram of a neural network combination model provided by a certain embodiment of the present invention;

[0133] Figure 6 A schematic structural diagram of a power grid load forecasting device based on multiple models provided by the present invention. DETAILED DESCRIPTION

[0134] 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.

[0135] 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.

[0136] 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.

[0137] Grid load forecasting generally follows the stages of data collection, preprocessing, forecast target determination, forecast model construction, data analysis and adjustment. Commonly used data for load forecasting include historical load data, meteorological data, calendar variables, demographic factors, socioeconomic data, and power system internal data. Commonly used load forecasting models include traditional forecasting models such as time series analysis models (such as ARIMA models), machine learning forecasting models (such as BP neural network models, LSTM long short-term memory recurrent neural networks, etc.), and deep learning forecasting models (such as MLP-based models). Power load forecasting is a dynamic and continuous process that requires continuous adjustment of forecasting models and methods based on new data and situations.

[0138] With the large-scale access of new energy, the load changes become more complex, increasing the difficulty of load forecasting, and the multi-level load is also a factor that is less considered in current load forecasting. In order to achieve high-accuracy load forecasting and establish an accurate multi-level relationship of power grid load, it is necessary to consider load factors, socio-economic factors and meteorological factors and build an effective load forecasting model.

[0139] like Figure 1 As shown, the present invention provides a power grid load forecasting method based on multiple models, which specifically includes the following steps:

[0140] Obtain historical load data of multiple substations to form initial historical load data, perform cluster analysis on the initial historical load data, and provide category division of the initial historical load data;

[0141] Through the improved LSTM model, based on the initial historical load data of each category, the load of each substation is predicted, and the load of the power network in the first area is predicted by combining the historical load data of the first area and the industrial data and meteorological data of the first area;

[0142] Based on the load forecast of the power network in the first area, the load forecast of the power network in the second area is performed by combining the historical load data of the second area with the constructed neural network combination model;

[0143] The first regional power network includes a plurality of substations, and the second regional power network includes a plurality of first regional power networks.

[0144] In a certain embodiment, the first regional power network is a municipal power network, and the second regional power network is a provincial power network.

[0145] That is, the load forecast for multi-level power grids considers the load forecast of substation-municipal-provincial power networks. For the load forecast of substations, taking the 220V substation as an example, the factors that need to be considered when conducting load forecasting include: load type (for example, commercial load, industrial load, residential load, agricultural load), attribute data of substations (for example, capacity-load ratio of substations, load rate, line structure of substations, and new energy output data). For the load forecast of municipal power networks, the factors that need to be considered include: load data of all districts or counties in the city, industrial data of the city, and meteorological data. For the load forecast of provincial power networks, the factors that need to be considered include: load data of all prefecture-level cities in the province, provincial industrial structure, meteorological data, etc.

[0146] For load forecasting of substations and municipal power networks, the available data is relatively direct and the data correlation is relatively strong. By improving the typical LSTM model and enhancing the dynamic prediction capability, the accuracy of power grid load forecasting can be improved.

[0147] Correspondingly, the load forecast of the provincial power network needs to rely on the load forecast data of the municipal power network, and also indirectly use the load forecast data of the substation. It has high requirements for the feature extraction of each data, and it is necessary to build a combined model to achieve accurate load forecasting for the provincial power network.

[0148] The historical load data of substations is relatively complex, and it is impossible to process the massive historical load data of each substation. The historical load data in the time series can be classified by clustering, and the grid load of the substation can be predicted in a targeted manner according to different types of historical load data.

[0149] like Figure 2 As shown, historical load data of multiple substations are obtained to form initial historical load data, and cluster analysis is performed on the initial historical load data to provide a type classification of the initial historical load data, including:

[0150] Preprocess the historical load data of multiple substations to form historical load curves of multiple time frames;

[0151] Resample each historical load curve to form a corresponding load sample;

[0152] By clustering multiple load samples, multiple groups of clusters are given;

[0153] Based on the association analysis of each group of clusters, multiple clusters are formed;

[0154] By comparing the historical load data of each substation with each cluster center, the historical load data of each substation is classified into the corresponding cluster to complete the type classification of the initial historical load data.

[0155] Resample each load curve to form a corresponding load sample, including:

[0156] Define the number of samples for resampling, the number of resampling times, and the statistical characteristics of resampling, perform statistical characteristic analysis on each resampled sample, fuse each statistical characteristic, perform deviation correction, set the confidence interval, and form the corresponding load sample;

[0157] Based on the association analysis of each group of clusters, multiple clusters are formed, including:

[0158] Calculate the centroid distance between each group of clusters, which is specifically expressed as:

[0159]

[0160] Among them, μ i ,μ j Cluster C i With cluster C j The center of mass, d(μ i ,μ j ) is cluster C i With cluster C j The distance between the centroids, μ ik ,μ jk They are μ in k dimension respectively. i and μ j The coordinates of the cluster, d is the dimension endpoint, n is the total number of clusters, x ik ,x jk Cluster C i With cluster C j Elements of

[0161] Based on the comparison with the threshold distance, clusters are merged to form cluster groups;

[0162] The cluster group is removed and updated into a new group of clusters, and the centroid distance calculation between each group of clusters is repeated until the association analysis of each group of clusters is completed, and the remaining groups of clusters correspond to each cluster.

[0163] For the clustering processing of historical load data of substations, hierarchical analysis can be used to overcome the problem that general massive data is easily affected by the K value and falls into local optimal solutions during the clustering process.

[0164] Multiple groups of clusters are determined by load samples, and then hierarchical association analysis is performed on the multiple groups of clusters. The hierarchical association analysis considers the centroid links of each group of clusters. The distance between the centroids of two clusters is calculated to determine the similarity between them, forming cluster groups, and finally obtaining clusters with consistent categories, which are used as characteristic factors for load prediction of each substation.

[0165] For load forecasting of substations and municipal power networks, the accuracy of grid load forecasting can be improved by improving the typical LSTM model and enhancing the dynamic forecasting capability. In addition, based on the hierarchical relationship between the municipal power network and each substation, in the load forecasting of the municipal power network, the load forecasting data of each corresponding substation can be directly used as a feature factor to perform load forecasting in the improved LSTM model.

[0166] like Figure 3 As shown in the figure, through the improved LSTM model, based on the initial historical load data of each category, the load forecast of each substation is carried out, and the load forecast of the power network in the first area is carried out in combination with the historical load data of the first area and the industrial data and meteorological data of the first area, which specifically includes:

[0167] By matching each substation with the initial historical load data of each category, combined with the substation attribute data and meteorological data, and using the improved LSTM model, the load forecast data of each substation is given;

[0168] The load forecast data of the substation, the historical load data of the first region, the industrial data of the first region, and the meteorological data are integrated, and the load forecast of the power network in the first region is carried out through the improved LSTM model.

[0169] like Figure 4 As shown in the figure, the LSTM model is a long short-term memory neural network. As a special network in the recurrent neural network, the internal storage unit mainly uses memory cells to replace traditional neurons. The typical LSTM model framework follows the gate mechanism, which consists of an input gate, a forget gate, and an output gate.

[0170] In the forget gate, through the action of the Sigmoid activation function, the hidden state h is output according to the time t-1 t-1 With input x at time t t Given the forget gate output f t .

[0171] In the input gate, the Sigmoid activation function is used to control the input x at time t. t , output hidden state h at time t-1 t-1 , and also use the tanh activation function to output the candidate LSTM to update.

[0172] In the output gate, adjust the output of the LSTM model and select the LSTM output c at time t-1 t-1 The valid information in is transferred to the LSTM at the next moment and outputted. The expression of the output gate is as follows:

[0173] o t =σ(W ho h t-1 +W xo x t +b o )

[0174] h t =o t *tanh(c t )

[0175] Among them, t is the output of the output gate after the Sigmoid activation function, * is the Schur product operation, σ() is the Sigmoid activation function, W ho , W xo is the weight group in the output gate after the Sigmoid activation function, b o is the bias parameter of the output gate when it passes through the Sigmoid activation function, h t-1 Output hidden state at time t-1, x t is the input at time t, h t is the hidden state output at time t, tanh() is the tanh activation function, c t is the LSTM output at time t.

[0176] The improved LSTM model includes adding a sigmoid neural network layer group to the LSTM model architecture, and the sigmoid neural network layer group includes a first sigmoid neural network layer, a second sigmoid neural network layer, and a fusion layer;

[0177] The first sigmoid neural network layer is used to control the input at the current moment and realize the Schur product calculation of the output value and the output hidden state at the previous moment to form the optimized output hidden state at the previous moment, which is specifically expressed as:

[0178] u t =σ(W u x t +b u )

[0179] h' t-1 =h t-1 *u t

[0180] Among them, u tis the output value at time t of the first sigmoid neural network layer, x t is the input at time t, W u is the weight of the first sigmoid neural network layer, b u is the bias parameter of the first sigmoid neural network layer, σ() is the Sigmoid activation function, and h t-1 Output hidden state at time t-1, h' t-1 is the optimized output hidden state at time t-1, * is the Schur product operation;

[0181] The second sigmoid neural network layer is used to control the optimized output hidden state of the previous moment, and realizes the calculation of the Schur product of the output value and the current moment input to form the optimized current moment input, which is specifically expressed as:

[0182] v t =σ(W v h t-1 '+b v )

[0183] x' t =x t *v t

[0184] Among them, v t is the output value at time t of the second sigmoid neural network layer, h t-1 Output hidden state at time t-1, W v is the weight of the second sigmoid neural network layer, b v is the bias parameter of the second sigmoid neural network layer, σ() is the Sigmoid activation function, x' t is the optimized input at time t, x t Input at time t;

[0185] The fusion layer is used to combine the weights and bias parameters to fuse the optimized hidden state of the previous moment output and the optimized current moment input, and then transmit them to the forget gate, input gate and output gate in the LSTM model architecture respectively. The specific expression is:

[0186] s t =(W 1 h' t-1 +b l1 )+(W 2 x' t +b l2 )

[0187] Among them, s t is the fusion output, W 1 , W2 are the weights of the first sigmoid neural network layer and the second sigmoid neural network layer in the fusion layer, b l1 , b l2 are the bias parameters corresponding to the first sigmoid neural network layer and the second sigmoid neural network layer in the fusion layer.

[0188] In a certain embodiment, the attribute data of the substation includes the capacity-load ratio, load rate and new energy output data of the substation;

[0189] By matching each substation with the initial historical load data of each category, combined with the attribute data and meteorological data of the substation, the load forecast data of each substation is given through the improved LSTM model, including:

[0190] The time series is determined by the initial historical load data of each substation, and the capacity-load ratio, load rate, new energy processing data, and meteorological data of the substation are processed according to the time series to form multi-dimensional feature data;

[0191] Sort out the input quantity, time series length, and feature dimension values ​​to form a three-dimensional input tensor;

[0192] The three-dimensional input tensor is fed into the improved LSTM model to give the load forecast data for each substation.

[0193] Similarly, the load forecast data of the substation, the historical load data of the first region, the industrial data of the first region, and the meteorological data are integrated to perform load forecasting on the power network of the first region through the improved LSTM model, including:

[0194] Determine a time series using the historical load data of the first region, and process the load forecast data of the corresponding power station, the industrial data of the first region, and the meteorological data according to the time series to form multi-dimensional feature data;

[0195] Sort out the input quantity, time series length, and feature dimension values ​​to form a three-dimensional input tensor;

[0196] The three-dimensional input tensor is fed into the improved LSTM model to give the load forecast data of the power network in the first region.

[0197] The above improvements on the typical LSTM model, the addition of the first sigmoid neural network layer, the second sigmoid neural network layer and the fusion layer, enhance the dynamic prediction capability and improve the accuracy of power grid load prediction at the substation level and the first regional power network (municipal level).

[0198] The load forecasting of the provincial power grid requires the use of the load forecasting data of the municipal power grid, and also indirectly uses the load forecasting data of the substation. It has high requirements for the feature extraction of each data and needs to be constructed through a combined model. Among them, LSTM is widely used in time series analysis due to its advantage of long-term memory of historical information. In addition, CNN has a strong feature extraction capability. Compared with a single model, the combination of the two is more conducive to improving the prediction accuracy.

[0199] At the same time, taking into account the complexity of characteristic factors in predicting the grid load of the second regional power network, the number of LSTM model groups of corresponding dimensions is set, and the corresponding number of LSTM model units are set in each group. The fusion of all characteristic factors is realized through the LSTM model group, and the intrinsic relationship between various characteristic factors and grid load is further explored from the output end, so as to improve the prediction effect of the grid load of the second regional power network (provincial power network).

[0200] like Figure 5 As shown in the figure, the construction of the neural network combination model specifically includes:

[0201] Multiple groups of LSTM models are used in parallel to build a corresponding multi-channel structure, where each group of LSTM models includes multiple LSTM model units;

[0202] The multi-channel structure is connected to the transposition layer, which is connected to the CNN model. The convolutional layer in the CNN model performs feature fusion and gives the prediction result.

[0203] Preferably, the LSTM model can be selected as follows Figure 4 The improved LSTM model shown in FIG. 1 includes adding a sigmoid neural network layer group to the LSTM model architecture, and the sigmoid neural network layer group includes a first sigmoid neural network layer, a second sigmoid neural network layer and a fusion layer.

[0204] Based on the load forecast of the power network in the first area, the load forecast of the power network in the second area is carried out by combining the historical load data of the second area through the constructed neural network combination model, which specifically includes:

[0205] Determine a time series using the historical load data of the second region, and process the load forecast data corresponding to the power network of the first region, the industrial data of the second region, and the meteorological data according to the time series to form multi-dimensional feature data, wherein the dimension of the feature data is the same as the number of groups of the LSTM model in the neural network combination model;

[0206] Input each dimension feature data into the LSTM model of the corresponding group respectively, and output the initial result sequence;

[0207] The initial structure sequence of each group of outputs is transposed and then input into the CNN model. After the feature fusion of the convolution layer, the load forecast data of the power network in the second area is given.

[0208] Among them, the feature fusion of the convolutional layer is specifically expressed as:

[0209]

[0210] Among them, S (1,n) is the output after feature fusion of the convolutional layer of the CNN model, f() is the activation function, b is the bias vector of the convolutional layer of the CNN model, M is the number of LSTM model groups, N is the number of LSTM model units in each group of LSTM models, and H s,p is the initial result output by the pth LSTM model unit in the sth group of LSTM models, ω s,p is the corresponding convolution kernel weight matrix.

[0211] In a certain embodiment, the load forecast data of the first regional power network is used as a dimensional feature data, represented as {X 1,1 ,X 1,2 ,…,X 1,N}, similarly, the industrial data and meteorological data of the second region as a dimensional feature data can also be represented in the form of a sequence, represented as {X 2,1 ,X 2,2 ,…,X 2,N}、{X 3,1 ,X 3,2 ,…,X 3,N Each of the above dimensional feature data sequences is input into the corresponding LSTM models for prediction, and the output initial result sequences are represented as H 1 =(H 1,1 ,H 1,2 ,…,H 1,N ), H 2 =(H 2,1 ,H 2,2 ,…,H 2,N ) and H 3 =(H 3,1 ,H 3,2 ,…,H 3,N ). Then, the above-mentioned initial output result sequence is fused through the convolution layer to explore the intrinsic relationship between various characteristic factors and grid load, so as to improve the prediction effect of grid load of the second regional power network (provincial power network).

[0212] like Figure 6 As shown, the present invention also provides a power grid load prediction device based on multiple models, specifically comprising:

[0213] The acquisition module is used to obtain historical load data of multiple substations, form initial historical load data, obtain historical load data of the first area, industrial data and meteorological data of the first area, and historical load data of the second area;

[0214] An analysis module is used to perform cluster analysis on the initial historical load data and provide a classification of the initial historical load data;

[0215] The prediction module is used to perform load prediction for each substation based on the initial historical load data of each category through the improved LSTM model, and to perform load prediction for the power network in the first area in combination with the historical load data of the first area and the industrial data and meteorological data of the first area; based on the load prediction of the power network in the first area, the load prediction for the power network in the second area is performed through the constructed neural network combination model in combination with the historical load data of the second area.

[0216] The present invention provides a multi-model-based power grid load forecasting method and device, which at least have the following beneficial effects:

[0217] (1) The present invention implements grid load prediction for multi-level power networks through multiple models in a bottom-up manner. Each model takes into account the characteristics of each level of the power network, which significantly improves the prediction accuracy of the grid load at each level.

[0218] (2) Multiple groups of clusters are determined based on load samples, and then hierarchical association analysis is performed on the multiple groups of clusters. The hierarchical association analysis considers the centroid links of each group of clusters and determines the similarity between them by calculating the distance between the centroids of two clusters to form cluster groups. Finally, clusters with consistent categories are obtained, which are used as characteristic factors for load prediction of each substation.

[0219] (3) By improving the typical LSTM model, adding the first sigmoid neural network layer, the second sigmoid neural network layer and the fusion layer, the dynamic prediction capability is enhanced, and the accuracy of power grid load prediction at the substation level and the first regional power network (municipal level) is improved.

[0220] (4) Taking into account the complexity of characteristic factors in the grid load prediction of the second regional power network, the number of LSTM model groups of corresponding dimensions is set, and a corresponding number of LSTM model units are set in each group. The fusion of all characteristic factors is realized through the LSTM model group, and the intrinsic relationship between various characteristic factors and grid load is further explored from the output end, thereby improving the prediction effect of the grid load of the second regional power network.

[0221] It should be noted that the computer-readable medium of the present disclosure may be a computer-readable signal medium or a computer-readable storage medium or any combination of the two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present disclosure, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, in which a computer-readable program code is carried. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The computer readable signal medium may also be any computer readable medium other than a computer readable storage medium, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.

[0222] The computer-readable medium may be included in the electronic device, or may exist independently without being installed in the electronic device.

[0223] Computer program code for performing the operations of the present disclosure may be written in one or more programming languages, or a combination thereof, including object-oriented programming languages, such as Java, Smalltalk, C++, and conventional procedural programming languages, such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0224] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or the flowchart, and the combination of boxes in the block diagram and / or the flowchart, can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0225] The units involved in the embodiments of the present disclosure may be implemented by software or hardware, wherein the name of a unit does not limit the unit itself in some cases.

[0226] 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-model based power grid load forecasting method, characterized in that: The specific steps include: Obtain historical load data of multiple substations to form initial historical load data, perform cluster analysis on the initial historical load data, and provide category division of the initial historical load data; Through the improved LSTM model, based on the initial historical load data of each category, the load of each substation is predicted, and the load of the power network in the first area is predicted by combining the historical load data of the first area and the industrial data and meteorological data of the first area; Based on the load forecast of the power network in the first area, the load forecast of the power network in the second area is performed by combining the historical load data of the second area with the constructed neural network combination model; The first regional power network includes a plurality of substations, and the second regional power network includes a plurality of first regional power networks.

2. The multi-model-based power grid load forecasting method according to claim 1, characterized in that: Obtain historical load data of multiple substations to form initial historical load data, perform cluster analysis on the initial historical load data, and provide the type classification of the initial historical load data, including: Preprocess the historical load data of multiple substations to form historical load curves of multiple time frames; Resample each historical load curve to form a corresponding load sample; By clustering multiple load samples, multiple groups of clusters are given; Based on the association analysis of each group of clusters, multiple clusters are formed; By comparing the historical load data of each substation with each cluster center, the historical load data of each substation is classified into the corresponding cluster to complete the type classification of the initial historical load data.

3. The multi-model-based power grid load forecasting method according to claim 2, characterized in that: Resample each load curve to form a corresponding load sample, including: Define the number of samples for resampling, the number of resampling times, and the statistical characteristics of resampling, perform statistical characteristic analysis on each resampled sample, fuse each statistical characteristic, perform deviation correction, set the confidence interval, and form the corresponding load sample; Based on the association analysis of each group of clusters, multiple clusters are formed, including: Calculate the centroid distance between each group of clusters, which is specifically expressed as: Among them, μ i ,μ j Cluster C i With cluster C j The center of mass, d(μ i ,μ j ) is cluster C i With cluster C j The distance between the centroids, μ ik ,μ jk They are μ in k dimension respectively. i and μ j The coordinates of the cluster, d is the dimension endpoint, n is the total number of clusters, x ik ,x jk Cluster C i With cluster C j Elements of Based on the comparison with the threshold distance, clusters are merged to form cluster groups; The cluster group is removed and updated into a new group of clusters, and the centroid distance calculation between each group of clusters is repeated until the association analysis of each group of clusters is completed, and the remaining groups of clusters correspond to each cluster.

4. The power grid load forecasting method based on multiple models as claimed in claim 1, characterized in that: Through the improved LSTM model, based on the initial historical load data of each category, the load forecast of each substation is carried out, and the load forecast of the power network in the first area is carried out in combination with the historical load data of the first area and the industrial data and meteorological data of the first area, including: By matching each substation with the initial historical load data of each category, combined with the substation attribute data and meteorological data, and using the improved LSTM model, the load forecast data of each substation is given; The load forecast data of the substation, the historical load data of the first region, the industrial data of the first region, and the meteorological data are integrated, and the load forecast of the power network in the first region is carried out through the improved LSTM model.

5. The multi-model-based power grid load forecasting method according to claim 4, characterized in that: The improved LSTM model includes adding a sigmoid neural network layer group to the LSTM model architecture, and the sigmoid neural network layer group includes a first sigmoid neural network layer, a second sigmoid neural network layer, and a fusion layer; The first sigmoid neural network layer is used to control the input at the current moment and realize the Schur product calculation of the output value and the output hidden state at the previous moment to form the optimized output hidden state at the previous moment; The second sigmoid neural network layer is used to control the optimized hidden state of the previous moment output and calculate the Schur product of the output value and the current moment input to form the optimized current moment input; The fusion layer is used to combine the weights and bias parameters to fuse the optimized hidden state of the previous moment output and the optimized input of the current moment, and then transmit them to the forget gate, input gate and output gate in the LSTM model architecture respectively.

6. The multi-model-based power grid load forecasting method according to claim 5, characterized in that: The attribute data of the substation includes the capacity-load ratio, load rate and renewable energy output data of the substation; By matching each substation with the initial historical load data of each category, combined with the attribute data and meteorological data of the substation, the load forecast data of each substation is given through the improved LSTM model, including: The time series is determined by the initial historical load data of each substation, and the capacity-load ratio, load rate, new energy processing data, and meteorological data of the substation are processed according to the time series to form multi-dimensional feature data; Sort out the input quantity, time series length, and feature dimension values ​​to form a three-dimensional input tensor; The three-dimensional input tensor is fed into the improved LSTM model to give the load forecast data for each substation.

7. The multi-model-based power grid load forecasting method according to claim 5, characterized in that: The load forecast data of the substation, the historical load data of the first region, the industrial data of the first region, and the meteorological data are integrated to perform load forecasting on the power network of the first region through the improved LSTM model, including: Determine a time series using the historical load data of the first region, and process the load forecast data of the corresponding power station, the industrial data of the first region, and the meteorological data according to the time series to form multi-dimensional feature data; Sort out the input quantity, time series length, and feature dimension values ​​to form a three-dimensional input tensor; The three-dimensional input tensor is fed into the improved LSTM model to give the load forecast data of the power network in the first region.

8. The power grid load forecasting method based on multiple models as claimed in claim 1, characterized in that: The construction of the neural network combination model includes: Multiple groups of LSTM models are used in parallel to build a corresponding multi-channel structure, where each group of LSTM models includes multiple LSTM model units; The multi-channel structure is connected to the transposition layer, which is connected to the CNN model. The convolutional layer in the CNN model performs feature fusion and gives the prediction result.

9. The multi-model-based power grid load forecasting method according to claim 8, characterized in that: Based on the load forecast of the power network in the first area, the load forecast of the power network in the second area is carried out by combining the historical load data of the second area through the constructed neural network combination model, which specifically includes: Determine a time series using the historical load data of the second region, and process the load forecast data corresponding to the power network of the first region, the industrial data of the second region, and the meteorological data according to the time series to form multi-dimensional feature data, wherein the dimension of the feature data is the same as the number of groups of the LSTM model in the neural network combination model; Input each dimension feature data into the LSTM model of the corresponding group respectively, and output the initial result sequence; The initial structure sequence of each group of outputs is transposed and then input into the CNN model. After the feature fusion of the convolution layer, the load forecast data of the power network in the second area is given. Among them, the feature fusion of the convolutional layer is specifically expressed as: Among them, S (1,n) is the output after feature fusion of the convolutional layer of the CNN model, f() is the activation function, b is the bias vector of the convolutional layer of the CNN model, M is the number of LSTM model groups, N is the number of LSTM model units in each group of LSTM models, and H s,p is the initial result output by the pth LSTM model unit in the sth group of LSTM models, ω s,p is the corresponding convolution kernel weight matrix.

10. A power grid load forecasting device based on multiple models, characterized in that: The method for predicting power grid load based on multiple models as claimed in any one of claims 1 to 9 specifically comprises: The acquisition module is used to obtain historical load data of multiple substations, form initial historical load data, obtain historical load data of the first area, industrial data and meteorological data of the first area, and historical load data of the second area; An analysis module is used to perform cluster analysis on the initial historical load data and provide a classification of the initial historical load data; The prediction module is used to perform load prediction for each substation based on the initial historical load data of each category through the improved LSTM model, and to perform load prediction for the power network in the first area in combination with the historical load data of the first area and the industrial data and meteorological data of the first area; based on the load prediction of the power network in the first area, the load prediction for the power network in the second area is performed through the constructed neural network combination model in combination with the historical load data of the second area.

Images