Power load prediction method and device, electronic equipment and medium

Abstract

The embodiment of the invention provides a power load prediction method and device, electronic equipment and a medium, and the method comprises the steps: obtaining the historical power load sequence data of a target power grid system, and enabling the historical power load sequence data to comprise power load information and temperature information; extracting a spatial feature vector and a time feature vector, the spatial feature vector being a multi-channel vector obtained by performing spatial feature extraction on the historical power load sequence data, and the time feature vector being a long and short term memory feature vector of at least one channel obtained by performing time feature extraction on the historical power load sequence data; performing feature fusion on the spatial feature vector and the time feature vector to generate a feature fusion vector; and predicting power load information of the target power grid system at a future time point based on the feature fusion vector. Therefore, power load data with complex characteristics can be processed, future power loads can be predicted more accurately, and safe and stable operation of a power system is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and particularly to a method, device, electronic device and medium for predicting electric power load. Background Art

[0002] Electric power load prediction is a process of predicting future electric power demand by using various data and methods. Electric power load prediction is of great significance in many aspects such as the planning, operation and scheduling of power systems. Its prediction results can provide an important basis for the future planning of power systems, including determining the installation of future generating units, the size, location and time of installed capacity, as well as the capacity expansion, reconstruction, construction and development of power grids.

[0003] Currently, electric power load prediction is mainly achieved through time series models or deep learning models in the field of artificial intelligence. Generally, by analyzing historical electric power load data, the law of its change over time is found, and a time series model, such as an Auto-Regressive Moving Average (ARMA) model, is established to predict future loads. However, since the change of load may be affected by various factors such as weather and economic conditions in addition to time, the time series model can only predict periodic characteristics and cannot cope with load fluctuations caused by emergencies, resulting in a decrease in prediction accuracy.

[0004] With the development of artificial intelligence technology, since deep learning models have strong non-linear mapping ability and learning ability to handle complex non-linear relationships. The Recurrent Neural Network (RNN) model performs relatively prominently in the time series field; among them, the Long Short-Term Memory (LSTM) model can focus on long-term dependencies in the sequence and use a gating mechanism to remember and transmit long-term information, so as to better handle long-term dependence problems and can effectively capture the long-term change trend of electric power load. However, the LSTM model updates the hidden state at each time step, and the transmission of information is a relatively smooth process, which may smooth out local and suddenly emerging key information, resulting in the LSTM model may ignore some local key information and affect the accuracy of prediction. Summary of the Invention

[0005] The present invention provides a method, device, electronic device and medium for predicting electric power load, so as to solve the defect of low accuracy of electric power load prediction in the prior art, realize more accurate electric power load prediction and improve prediction accuracy.

[0006] The present invention provides a method for predicting electric power load, including: Obtain the historical power load sequence data of the target power grid system, where the historical power load sequence data includes power load information and temperature information; Extract the spatial feature vector and the temporal feature vector of the historical power load sequence data. The spatial feature vector is a multi-channel vector obtained by performing spatial feature extraction on the historical power load sequence data, and the temporal feature vector is at least one channel of long short-term memory feature vectors obtained by performing temporal feature extraction on the historical power load sequence data; Fuse the spatial feature vector and the temporal feature vector to generate a feature fusion vector; Predict the power load information at future time points of the target power grid system based on the feature fusion vector.

[0007] In a possible implementation manner, the method further includes: Collect the historical power load data corresponding to multiple historical time points of the target power grid system based on a preset time period to obtain the historical power load sequence data.

[0008] In a possible implementation manner, the method further includes: Perform spatial feature extraction on the historical power load sequence data through a convolutional neural network to obtain a multi-channel spatial feature vector; Perform temporal feature extraction on the historical power load sequence data through a long short-term memory network to obtain at least one channel of temporal feature vectors, and the temporal feature vectors are long short-term memory feature vectors.

[0009] In a possible implementation manner, the method further includes: The convolutional neural network includes a plurality of serially arranged convolutional units, and the convolutional unit includes a first convolutional layer, a second convolutional layer, and a third convolutional layer; When there is an (i - 2)-th initial eigenvalue, perform convolutional processing on the i-th initial eigenvalue, the (i - 1)-th initial eigenvalue, and the (i - 2)-th initial eigenvalue in the initial sequence data through the first convolutional layer to generate the i-th first eigenvalue. The initial sequence data includes a plurality of initial eigenvalues, and the initial sequence data input to the first convolutional layer of the first convolutional unit is the historical power load sequence data; When there is no (i - 2)-th initial eigenvalue, perform convolutional processing on the i-th initial eigenvalue and the (i - 1)-th initial eigenvalue in the initial sequence data through the first convolutional layer to generate the i-th first eigenvalue, where i ≥ 2; Input the first sequence data including each first eigenvalue into the second convolutional layer; When there is a (j - 4)-th first eigenvalue, the second convolutional layer performs convolutional processing on the j-th first eigenvalue, the (j - 2)-th first eigenvalue, and the (j - 4)-th first eigenvalue in the first sequence data to generate a j-th second eigenvalue; When there is no (j - 4)-th first eigenvalue, the second convolutional layer performs convolutional processing on the j-th first eigenvalue and the (j - 2)-th first eigenvalue in the first sequence data to generate a j-th second eigenvalue, where j≥4; Input the second sequence data containing each second eigenvalue into the third convolutional layer; When there is a (k - 6)-th second eigenvalue, the third convolutional layer performs convolutional processing on the k-th second eigenvalue, the (k - 3)-th second eigenvalue, and the (k - 6)-th second eigenvalue in the second sequence data to generate a k-th third eigenvalue; When there is no (k - 6)-th second eigenvalue, the third convolutional layer performs convolutional processing on the k-th second eigenvalue and the (k - 3)-th second eigenvalue in the second sequence data to generate a k-th third eigenvalue, obtaining a third sequence data of each third eigenvalue, where k≥7; Based on the third sequence data output by the last convolutional unit, obtain a multi-channel spatial feature vector.

[0010] In a possible implementation manner, the method further includes: The feature fusion network splices the spatial feature vector and the temporal feature vector along the channel dimension to generate a feature fusion vector.

[0011] In a possible implementation manner, the method further includes: Perform feature extraction on the feature fusion vector in the time dimension to obtain an intermediate fusion vector; Perform feature extraction on the intermediate fusion vector in the channel dimension to obtain a target fusion vector; Predict the power load information of the target power grid system at a future time point according to the target fusion vector.

[0012] In a possible implementation manner, the method further includes: Perform transpose processing on the feature fusion vector to obtain a first feature vector of C×T; Perform feature extraction on the first feature vector row by row to obtain a second feature vector of C×T; Perform transpose processing on the second feature vector to obtain a third feature vector of T×C, and generate the intermediate fusion vector according to the third feature vector; Extract features from the intermediate fusion vector row by row to generate a fourth feature vector of T×C, and generate the target fusion vector according to the fourth feature vector; The size of the feature fusion vector is T×C, and C = c1 + c2, where c1 is the number of channels of the spatial feature vector, c1≥2, c2 is the number of channels of the temporal feature vector, c2≥1, and T is the number of eigenvalues of the spatial feature vector and the temporal feature vector.

[0013] The present invention also provides a power load prediction device, including the following modules: A data acquisition module for acquiring historical power load sequence data of a target power grid system, where the historical power load sequence data includes power load information and temperature information; A feature extraction module for extracting a spatial feature vector and a temporal feature vector from the historical power load sequence data. The spatial feature vector is a multi-channel vector obtained by performing spatial feature extraction on the historical power load sequence data, and the temporal feature vector is at least one-channel long short-term memory feature vector obtained by performing temporal feature extraction on the historical power load sequence data; A feature fusion module for fusing the spatial feature vector and the temporal feature vector to generate a feature fusion vector; A prediction module for predicting the power load information at a future time point of the target power grid system based on the feature fusion vector.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the power load prediction method as described in any one of the above.

[0015] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the power load prediction method as described in any one of the above.

[0016] The present invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the power load prediction method as described in any one of the above.

[0017] The power load prediction method, device, electronic device and medium provided by the present invention obtain historical power load sequence data of a target power grid system, where the historical power load sequence data includes power load information and temperature information; extract spatial feature vectors and temporal feature vectors of the historical power load sequence data, the spatial feature vectors are multi-channel vectors obtained by performing spatial feature extraction on the historical power load sequence data, and the temporal feature vectors are at least one-channel long short-term memory feature vectors obtained by performing temporal feature extraction on the historical power load sequence data; fuse the spatial feature vectors and the temporal feature vectors to generate a feature fusion vector; and predict the power load information at a future time point of the target power grid system based on the feature fusion vector. Compared with the existing power load prediction methods using time series models or deep learning models in the field of artificial intelligence, which are affected by various factors such as weather and economic conditions, or have low prediction accuracy due to ignoring some local key information, in this solution, a convolutional neural network and a long short-term memory network are combined to perform feature extraction in different dimensions, and finally the features in different dimensions are fused to generate a feature fusion vector containing multi-dimensional features. Based on this feature fusion vector for power load prediction, it can process power load data with complex characteristics, can more accurately predict future power loads, and provide a strong guarantee for the safe and stable operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0019] Figure 1 is one of the flow diagrams of the power load prediction method provided by the present invention.

[0020] Figure 2 is the second flow diagram of the power load prediction method provided by the present invention.

[0021] Figure 3 is the structural diagram of the load prediction model provided by the present invention.

[0022] Figure 4 is the structural diagram of the convolutional neural network provided by the present invention.

[0023] Figure 5 is the schematic diagram of the principle of the convolutional unit provided by the present invention.

[0024] Figure 6It is a schematic structural diagram of the feature fusion network provided by the present invention.

[0025] Figure 7 It is a schematic diagram of the prediction layer generating the target fusion vector provided by the present invention.

[0026] Figure 8 It is a schematic structural diagram of the power load prediction device provided by the present invention.

[0027] Figure 9 It is a schematic structural diagram of the electronic device provided by the present invention. Detailed implementation manners

[0028] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Apparently, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without making creative efforts shall fall within the protection scope of the present invention.

[0029] To facilitate the understanding of the embodiments of the present invention, the following will further explain and illustrate with specific embodiments in conjunction with the accompanying drawings. The embodiments do not constitute a limitation to the embodiments of the present invention.

[0030] Figure 1 It is one of the schematic flowcharts of the power load prediction method provided by the present invention. As Figure 1 shown, the method includes the following: S11. Obtain the historical power load sequence data of the target power grid system.

[0031] In the embodiments of the present invention, first, the historical sequence data of the power load of the target power grid system is obtained. The historical sequence data includes historical feature values at multiple historical time points, and the historical feature values include power load values.

[0032] Specifically, based on a preset time period, the historical power load data corresponding to multiple historical time points of the target power grid system can be collected to obtain the historical power load sequence data, where the historical power load data includes power load information and temperature information.

[0033] S12. Extract the spatial feature vector and the time feature vector of the historical power load sequence data.

[0034] The spatial feature extraction of the historical power load sequence data is performed through a convolutional neural network to obtain a multi-channel spatial feature vector; the time feature extraction of the historical power load sequence data is performed through a long short-term memory network to obtain at least one channel of time feature vectors, and the time feature vectors are long short-term memory feature vectors.

[0035] Specifically, when there is an (i - 2)-th initial eigenvalue, the first convolutional layer performs convolutional processing on the i-th initial eigenvalue, the (i - 1)-th initial eigenvalue, and the (i - 2)-th initial eigenvalue in the initial sequence data to generate the i-th first eigenvalue. The initial sequence data includes multiple initial eigenvalues, and the initial sequence data input to the first convolutional layer of the first convolutional unit is historical power load sequence data; when there is no (i - 2)-th initial eigenvalue, the first convolutional layer performs convolutional processing on the i-th initial eigenvalue and the (i - 1)-th initial eigenvalue in the initial sequence data to generate the i-th first eigenvalue, where i ≥ 2; the first sequence data containing each first eigenvalue is input to the second convolutional layer; when there is a (j - 4)-th first eigenvalue, the second convolutional layer performs convolutional processing on the j-th first eigenvalue, the (j - 2)-th first eigenvalue, and the (j - 4)-th first eigenvalue in the first sequence data to generate the j-th second eigenvalue; when there is no (j - 4)-th first eigenvalue, the second convolutional layer performs convolutional processing on the j-th first eigenvalue and the (j - 2)-th first eigenvalue in the first sequence data to generate the j-th second eigenvalue, where j ≥ 4; the second sequence data containing each second eigenvalue is input to the third convolutional layer; when there is a (k - 6)-th second eigenvalue, the third convolutional layer performs convolutional processing on the k-th second eigenvalue, the (k - 3)-th second eigenvalue, and the (k - 6)-th second eigenvalue in the second sequence data to generate the k-th third eigenvalue; when there is no (k - 6)-th second eigenvalue, the third convolutional layer performs convolutional processing on the k-th second eigenvalue and the (k - 3)-th second eigenvalue in the second sequence data to generate the k-th third eigenvalue, obtaining the third sequence data of each third eigenvalue, where k ≥ 7; a multi-channel spatial feature vector is obtained based on the third sequence data output by the last convolutional unit.

[0036] The long short-term memory network predicts the time-point eigenvalues corresponding to each time step of the historical power load sequence data to obtain the hidden state of the next time step; based on the hidden state, the feature vectors of multiple time steps are determined to obtain the time feature vectors of at least one channel.

[0037] S13. Feature-fuse the spatial feature vector and the time feature vector to generate a feature-fused vector.

[0038] The feature-fusion network concatenates the spatial feature vector and the time feature vector along the channel dimension to generate a feature-fused vector.

[0039] Specifically, the feature fusion network includes a fusion layer and a prediction layer. The fusion layer is used to fuse the spatial feature vector (a convolutional feature vector) and the temporal feature vector (a long short-term memory feature vector) in the channel dimension to generate a feature fusion vector. The fusion layer is used to fuse the convolutional feature vector and the long short-term memory feature vector in the channel dimension to generate a feature fusion vector.

[0040] S14. Predict the power load information of the target power grid system at a future time point based on the feature fusion vector.

[0041] Perform feature extraction on the feature fusion vector in the time dimension to obtain an intermediate fusion vector; perform feature extraction on the intermediate fusion vector in the channel dimension to obtain a target fusion vector; predict the power load information of the target power grid system at a future time point based on the target fusion vector.

[0042] The power load prediction method provided by the present invention obtains the historical power load sequence data of the target power grid system, wherein the historical power load sequence data includes power load information and temperature information; extracts the spatial feature vector and the temporal feature vector of the historical power load sequence data. The spatial feature vector is a multi-channel vector obtained by performing spatial feature extraction on the historical power load sequence data, and the temporal feature vector is at least one-channel long short-term memory feature vector obtained by performing temporal feature extraction on the historical power load sequence data; fuses the spatial feature vector and the temporal feature vector to generate a feature fusion vector; predicts the power load information of the target power grid system at a future time point based on the feature fusion vector. Compared with the existing power load prediction methods using time series models or deep learning models in the field of artificial intelligence, which are affected by various factors such as weather and economic conditions, or have low prediction accuracy due to ignoring some local key information, this method combines a convolutional neural network and a long short-term memory network to perform feature extraction in different dimensions respectively, and finally fuses the features in different dimensions to generate a feature fusion vector containing multi-dimensional features. Based on this feature fusion vector for power load prediction, it can process power load data with complex characteristics, can more accurately predict future power loads, and provides a strong guarantee for the safe and stable operation of the power system.

[0043] Figure 2 is the second schematic diagram of the flow of the power load prediction method provided by the present invention, as Figure 2 shown, this method includes the following: S21. Based on a preset time period, collect the historical power load data corresponding to multiple historical time points of the target power grid system to obtain historical power load sequence data.

[0044] In the embodiments of the present invention, the historical power load data includes power load information and temperature information.

[0045] During the operation of power supply equipment such as power grid systems or generators, the power load value at corresponding time points can be collected at regular intervals. This power load value is used to represent the magnitude of the power load, and specifically, it can be the load power. For example, if the collection is performed every 15 minutes, a total of 96 power load values at different time points can be collected within a day (24 hours).

[0046] For each sampled time point, the historical feature value including the corresponding power load value can be determined. By combining the historical feature values of multiple consecutive time points, the corresponding sequence data can be formed. Since this sequence data is obtained based on the already collected data, it is called historical sequence data. Correspondingly, the time points for collecting power load values are called historical time points.

[0047] Optionally, the historical feature value of each time point can further include: the time information corresponding to the historical time point, and the temperature information at the historical time point.

[0048] In this embodiment, the power load generally has a strong periodic pattern. Considering the time information as a kind of historical feature value is more conducive to extracting features in the time dimension. Moreover, many power loads are also related to the actual environmental temperature. For example, in winter or summer, the power load is generally higher than that in spring or autumn. Considering the temperature information as a kind of historical feature value enables the load prediction model to learn the correlation between temperature and power load, and can perform more reasonable load prediction when encountering situations such as sudden temperature changes.

[0049] It should be noted that for each historical time point, its historical feature value can specifically include: the power load value, time information, and temperature information at this historical time point.

[0050] S22. Extract spatial features from the historical power load sequence data through a convolutional neural network to obtain a multi-channel spatial feature vector.

[0051] S23. Extract time features from the historical power load sequence data through a long short-term memory network to obtain at least one channel of time feature vectors, and the time feature vectors are long short-term memory feature vectors.

[0052] In the embodiments of the present invention, a power load prediction model is pre-constructed and trained. When load prediction needs to be performed based on historical sequence data, the historical sequence data is input into the power load prediction model. After the model processes the historical sequence data, the power load value at the future time point after the historical sequence data can be output.

[0053] Among them, the future time point can be the next time point after the last historical time point in the historical sequence data. Moreover, it is possible to predict the power load value at only one future time point, or to predict the power load values at multiple future time points. This embodiment does not limit this. During model training, for a piece of historical data, after prediction in the training stage, the predicted load value L1 is output, and the future actual load value L2 is used as the data label. A loss function is constructed based on L1 and L2 and minimized. The most commonly used loss function is to calculate the quadratic norm of L1 - L2. Generally, conventional methods such as gradient descent are used to train the model. After the loss function is minimized, the trained power load prediction model is obtained.

[0054] As Figure 3 shown, the load prediction model includes a convolutional neural network 201, a long short-term memory network 202, and a feature fusion network 203.

[0055] The convolutional neural network 201 is used to extract spatial features from the historical sequence data to determine a multi-channel convolutional feature vector. The long short-term memory network 202 is used to extract temporal features from the historical sequence data to determine at least one-channel long short-term memory feature vector. The feature fusion network 203 is used to fuse the convolutional feature vector and the long short-term memory feature vector in the channel dimension to generate a feature fusion vector; and predict the power load value at the future time point according to the feature fusion vector.

[0056] In this embodiment, the historical sequence data is used as the input data of the convolutional neural network 201 and the long short-term memory network 202. Feature extraction in different dimensions is respectively performed on the historical sequence data based on the convolutional neural network 201 and the long short-term memory network 202, so as to form feature vectors in different dimensions.

[0057] Specifically, convolutional processing is performed on the historical sequence data based on the convolutional neural network 201, so that feature extraction can be realized, and the corresponding feature vector, that is, the convolutional feature vector, can be determined. Generally, the convolutional neural network 201 is provided with multiple convolutional layers, and through multi-layer convolutional processing, the features in the historical sequence data can be extracted more comprehensively.

[0058] The long short-term memory network 202 can adopt an existing network architecture. For the historical sequence data containing the feature values of multiple historical time points, the feature values of the corresponding historical time points are respectively predicted at each time step to obtain the hidden state of the next time step. Based on these hidden states, the feature vectors of multiple time steps, that is, the long short-term memory feature vectors, can be determined.

[0059] For example, if the historical sequence data corresponds to N historical time points, that is, it contains N historical feature values, after the long short-term memory network 202 is processed through N time steps, N hidden states can be obtained, thereby generating an N-dimensional long short-term memory feature vector.

[0060] Since the long short-term memory network 202 mainly extracts features in the time dimension, the feature extraction process implemented based on the long short-term memory network 202 is called time feature extraction. And the convolution processing operation in the convolutional neural network 201 can mainly be used for static features in the historical sequence data. For the convenience of distinction, the feature extraction process implemented based on the convolutional neural network 201 is called spatial feature extraction.

[0061] In addition, when performing spatial feature extraction based on the convolutional neural network 201, generally, feature extraction is performed based on multiple convolutional kernels, so that the historical sequence data can be converted into a multi-channel (Channel) convolutional feature vector. When performing time feature extraction based on the long short-term memory network 202, a long short-term memory feature vector with a corresponding number of channels can be generated based on the number of LSTM units in the long short-term memory network 202, and the long short-term memory feature vector is a feature vector with at least one channel.

[0062] For example, if the number of channels of the convolutional feature vector is c1 and the number of channels of the long short-term memory feature vector is c2, then c1≥2 and c2≥1. Generally, multiple LSTM units can be set in the long short-term memory network 202, that is, c2≥2.

[0063] As Figure 3 shown, for the convolutional feature vector and the long short-term memory feature vector respectively determined by the convolutional neural network 201 and the long short-term memory network 202, the two can be fused based on the feature fusion network 203 to obtain a fused feature vector, that is, a feature fusion vector. Among them, the feature fusion network 203 fuses the two feature vectors (that is, the convolutional feature vector and the long short-term memory feature vector) in the channel dimension. For example, the two feature vectors are concatenated along the channel dimension, and the fusion of the feature vectors can be realized without loss of information.

[0064] After determining the feature fusion vector containing time features and spatial features, load prediction is performed based on this feature fusion vector, and the power load value at a future time point can be obtained more accurately. For example, the load can be predicted based on other LSTM network structures for this feature fusion vector, or the prediction can be realized based on a fully connected layer. This embodiment does not make a limitation on this.

[0065] As Figure 4 shown, the convolutional neural network 201 includes a plurality of serially arranged convolutional units 301; the convolutional unit 301 includes a first convolutional layer, a second convolutional layer, and a third convolutional layer.

[0066] The first convolutional layer is used to perform convolutional processing on the i-th initial eigenvalue, the (i - 1)-th initial eigenvalue, and the (i - 2)-th initial eigenvalue in the initial sequence data to generate the i-th first eigenvalue when the (i - 2)-th initial eigenvalue exists; when the (i - 2)-th initial eigenvalue does not exist, perform convolutional processing on the i-th initial eigenvalue and the (i - 1)-th initial eigenvalue in the initial sequence data to generate the i-th first eigenvalue, where i ≥ 2. Input the first sequence data containing each first eigenvalue into the second convolutional layer; wherein, the initial sequence data includes a plurality of initial eigenvalues, and the initial sequence data input to the first convolutional unit 301 is historical sequence data. That is, the initial eigenvalues input to the first convolutional unit 301 are the above-mentioned historical eigenvalues.

[0067] The second convolutional layer is used to perform convolutional processing on the j-th first eigenvalue, the (j - 2)-th first eigenvalue, and the (j - 4)-th first eigenvalue in the first sequence data to generate the j-th second eigenvalue when the (j - 4)-th first eigenvalue exists; when the (j - 4)-th first eigenvalue does not exist, perform convolutional processing on the j-th first eigenvalue and the (j - 2)-th first eigenvalue in the first sequence data to generate the j-th second eigenvalue, where j ≥ 4; input the second sequence data containing each second eigenvalue into the third convolutional layer.

[0068] The third convolutional layer is used to perform convolutional processing on the k-th second eigenvalue, the (k - 3)-th second eigenvalue, and the (k - 6)-th second eigenvalue in the second sequence data to generate the k-th third eigenvalue when the (k - 6)-th second eigenvalue exists; when the (k - 6)-th second eigenvalue does not exist, perform convolutional processing on the k-th second eigenvalue and the (k - 3)-th second eigenvalue in the second sequence data to generate the k-th third eigenvalue, obtaining the third sequence data of each third eigenvalue, where k ≥ 7; the third sequence data containing each third eigenvalue is the output data of the convolutional unit 301.

[0069] Among them, the number of eigenvalues in the initial sequence data, the first sequence data, the second sequence data, and the third sequence data are respectively: m, m - 1, m - 3, m - 6. That is, m is the number of initial eigenvalues in the initial sequence data.

[0070] Figure 5 The schematic diagram of the principle of a convolutional unit 301 is shown, as Figure 5 shown, the initial sequence data includes 10 initial eigenvalues (i.e., m = 10), which are a1 to a10 respectively.

[0071] The first convolutional layer can obtain corresponding first eigenvalues and generate first sequence data by performing convolutional processing on three consecutive initial eigenvalues (i.e., the i-th initial eigenvalue, the i - 1-th initial eigenvalue, and the i - 2-th initial eigenvalue). As Figure 5 shown, the first sequence data includes 9 first eigenvalues (i.e., m - 1 = 9), which are b2 to b10 respectively.

[0072] The second convolutional layer can obtain corresponding second eigenvalues and generate second sequence data by performing convolutional processing on three first eigenvalues with an interval of 1 eigenvalue in sequence (i.e., the j-th first eigenvalue, the j - 2-th first eigenvalue, and the j - 4-th first eigenvalue). As Figure 5 shown, the second sequence data includes 7 second eigenvalues (i.e., m - 3 = 7), which are d4 to d10 respectively.

[0073] The third convolutional layer can obtain corresponding third eigenvalues and generate third sequence data by performing convolutional processing on three second eigenvalues with an interval of 2 eigenvalues in sequence (i.e., the k-th second eigenvalue, the k - 3-th second eigenvalue, and the k - 6-th second eigenvalue). As Figure 5 shown, the third sequence data includes 4 third eigenvalues (i.e., m - 6 = 4), which are h7 to h10 respectively.

[0074] In this embodiment, when the three convolutional layers (i.e., the first convolutional layer, the second convolutional layer, and the third convolutional layer) perform convolutional processing, the input eigenvalues have gradually increasing intervals, so that the finally generated third eigenvalues are related to more initial eigenvalues; as Figure 5 shown, the third eigenvalue h10 is obtained from the initial eigenvalues a1 to a10. Each initial eigenvalue corresponds to the power load value at a historical time point, which is convenient for load prediction based on data at more time points. Moreover, the finally determined third eigenvalue is also only related to part of the data in the initial sequence data, that is, the convolutional unit 301 is mainly used to extract local features in the initial sequence data, and it is easier to learn the local mutation information in the historical sequence data (the initial sequence data of the first convolutional unit 301 is the historical sequence data), and the load prediction can be realized more accurately.

[0075] Optionally, as Figure 4 shown, the convolutional neural network 201 further includes a screening layer 302 and a one-dimensional convolutional layer 303.

[0076] The screening layer 302 is used to remove the oldest 6n eigenvalues in the historical sequence data and generate intermediate sequence data. The one-dimensional convolutional layer 303 is used to perform one-dimensional convolutional processing on the intermediate sequence data to generate one-dimensional sequence data containing multiple eigenvalues. Among them, the one-dimensional sequence data and the output data of the last convolutional unit 301 are added to obtain a convolutional feature vector.

[0077] As Figure 5 shown, for each convolutional unit 301, after being processed by three convolutional layers, compared with the initial sequence data at the beginning, the third sequence data has 6 oldest eigenvalues less; and since the convolutional neural network 201 has n convolutional units 301, after being processed by each convolutional unit 301, 6 oldest eigenvalues will be reduced. Therefore, after being processed by n convolutional units 301 for the historical sequence data, the generated output data (i.e., the output data of the last convolutional unit 301) has 6n oldest eigenvalues less.

[0078] In this embodiment, as Figure 4 shown, a residual structure is introduced into the convolutional neural network 201 to avoid problems such as vanishing gradients. In the residual branch, a screening layer 302 is introduced to remove the 6n oldest eigenvalues from the historical sequence data, so that the output data of the residual branch at the end has the same size as the output data of the last convolutional unit 301. Moreover, for the intermediate sequence data generated by the screening layer 302, one-dimensional convolutional processing is performed on it based on the one-dimensional convolutional layer 303, and one-dimensional sequence data containing multiple eigenvalues can be generated. Finally, by adding the one-dimensional sequence data and the output data of the last convolutional unit 301, the final output data of the convolutional neural network 201, i.e., the convolutional feature vector, can be obtained.

[0079] Optionally, the size of the convolutional feature vector is T×c1, and the size of the long short-term memory feature vector is T×c2. Wherein, c1 is the number of channels of the convolutional feature vector, c1≥2; c2 is the number of channels of the long short-term memory feature vector, c2≥1; T is the number of eigenvalues of the convolutional feature vector and the long short-term memory feature vector.

[0080] In this embodiment, the number of eigenvalues of the convolutional feature vector and the long short-term memory feature vector is the same, both being T. For example, T can be the number of historical eigenvalues in the historical sequence data. Or, if as Figure 3 and Figure 4 shown, the generated convolutional feature vector has 6n eigenvalues less than the historical sequence data, so T is 6n less than the number of historical eigenvalues, or rather, the number of historical eigenvalues in the historical sequence data is T + 6n.

[0081] Setting the convolutional feature vector and the long short-term memory feature vector to have the same number of eigenvalues T facilitates subsequent feature fusion in the channel dimension.

[0082] S24. Concatenate the spatial feature vector and the temporal feature vector along the channel dimension through a feature fusion network to generate a feature fusion vector.

[0083] S25. Extract features from the feature fusion vector in the time dimension to obtain an intermediate fusion vector.

[0084] S26. Extract features from the intermediate fusion vector in the channel dimension to obtain a target fusion vector.

[0085] S27. Predict the power load information at a future time point of the target power grid system according to the target fusion vector.

[0086] As Figure 6 shown, the feature fusion network 203 includes: a fusion layer 501 and a prediction layer 502.

[0087] The fusion layer 501 is used to fuse the convolutional feature vector and the long short-term memory feature vector in the channel dimension to generate a feature fusion vector; the size of the feature fusion vector is T×C, and C = c1 + c2.

[0088] The prediction layer 502 is used to first extract features from the feature fusion vector in the time dimension to determine an intermediate fusion vector, and then extract features from the intermediate fusion vector in the channel dimension to determine a target fusion vector; predict the power load value at a future time point according to the target fusion vector.

[0089] In this embodiment, for the convolutional feature vector of T×c1 and the long short-term memory feature vector of T×c2, the fusion layer 501 splices the two in the channel dimension, so as to generate a feature vector of T×C (where C = c1 + c2), that is, the feature fusion vector. Moreover, based on the processing idea of depthwise separable convolution, this embodiment sequentially extracts features from the feature fusion vector in the time dimension and in the channel dimension, and finally obtains a fusion vector with strong fusion of the convolutional feature vector and the long short-term memory feature vector, that is, the target fusion vector. Subsequently, load prediction is performed based on this target fusion vector, and the power load value at a future time point can be determined more accurately.

[0090] Optionally, Figure 7 shows a schematic diagram of the prediction layer 502 generating a target fusion vector. The prediction layer 502 includes: a first transposed layer 5021, a first feature extraction layer 5022, a second transposed layer 5023, and a second feature extraction layer 5024.

[0091] The first transposed layer 5021 is used to transpose the feature fusion vector to obtain a first feature vector of C×T.

[0092] The first feature extraction layer 5022 is used to extract features from the first feature vector row by row to obtain a second feature vector of C×T.

[0093] The second transpose layer 5023 is used to transpose the second feature vector to obtain a third feature vector of T×C; and generate an intermediate fusion vector according to the third feature vector.

[0094] The second feature extraction layer 5024 is used to extract features from the intermediate fusion vector row by row to generate a fourth feature vector of T×C; and generate a target fusion vector according to the fourth feature vector.

[0095] Among them, when the prediction layer 502 extracts features, a residual structure can also be introduced; as Figure 7 shown, the intermediate fusion vector is obtained by adding the third feature vector and the feature fusion vector; the target fusion vector is obtained by adding the fourth feature vector and the intermediate fusion vector.

[0096] The feature fusion vector is a feature vector of T×C. For ease of description, different shading is used to represent the feature values at different time points, and different channels are divided by dotted lines. Figure 7 Taking T = 6 and C = 4 as an example. After the transpose processing of the first transpose layer 5021, the feature fusion vector of T×C is transposed into a feature vector of C×T, that is, the first feature vector.

[0097] At this time, each row of the first feature vector corresponds to the corresponding channels. The first feature extraction layer 5022 extracts features from the first feature vector row by row, so that the features in the time dimension corresponding to each channel can be extracted, that is, the feature extraction in the time dimension is realized, and thus a second feature vector of C×T is obtained; then, based on the second transpose layer 5023, transpose processing is performed again, and finally a feature vector of T×C is restored, that is, the third feature vector, which has the same size as the feature fusion vector.

[0098] Based on the first residual structure, adding the third feature vector of T×C and the feature fusion vector of T×C can obtain an intermediate fusion vector that extracts features in the time dimension.

[0099] At this time, each row of the intermediate fusion vector corresponds to each time point. Based on the second feature extraction layer 5024, extracting features from the intermediate fusion vector row by row again, the features in the channel dimension corresponding to each time point can be extracted, that is, the feature extraction in the channel dimension is realized, and thus a fourth feature vector of T×C is obtained. Finally, based on the first residual structure, adding the fourth feature vector and the intermediate fusion vector, the required target fusion vector is obtained. Subsequently, the load prediction can be performed on the target fusion vector based on the fully connected layer to determine the power load value at the future time point.

[0100] Among them, the first feature extraction layer 5022 and the second feature extraction layer 5024 can be implemented based on a feedforward neural network, and feature extraction is realized through multiple hidden layers.

[0101] ; 。

[0102] Among them, is the first eigenvector, is the second eigenvector; , are two weight matrices in the first feature extraction layer; , are two bias matrices in the first feature extraction layer; is the intermediate fusion vector, is the fourth eigenvector, , are two weight matrices in the second feature extraction layer; , are two bias matrices in the second feature extraction layer; is the activation function, and 。

[0103] In this embodiment, both the first feature extraction layer 5022 and the second feature extraction layer 5024 are provided with two hidden layers, and each hidden layer corresponds to a corresponding weight matrix and bias matrix. By training the feature fusion network 203, the weight matrices and bias matrices of each hidden layer can be determined. Specifically, the weight matrix and the bias matrix of the first hidden layer in the first feature extraction layer can be determined, as well as the weight matrix and the bias matrix of the second hidden layer. Similarly, the weight matrix and the bias matrix of the first hidden layer in the second feature extraction layer can be determined, as well as the weight matrix and the bias matrix of the second hidden layer.

[0104] And, an activation function is set between the two hidden layers to improve the prediction ability of the model for non-linear characteristics. Among them, some activation functions are not differentiable at certain positions, so this embodiment sets the activation function based on the probability density function of the normal distribution.

[0105] Specifically, the activation function is: , and after simplifying it, we get: 。 The activation function is differentiable and has a certain continuity.

[0106] The power load prediction method provided by the embodiment of the present invention combines a convolutional neural network and a long short-term memory network to perform feature extraction in the spatial dimension and the time dimension respectively, and fuse the features of different dimensions, so as to generate a feature fusion vector containing multi-dimensional features. Through the processing of three convolutional layers of the convolutional unit, local features in the historical sequence data can be extracted, global feature information can be strengthened, and combined with the long short-term memory feature vector extracted by the long short-term memory network, the characteristics of the historical sequence data can be determined more comprehensively and accurately, so as to perform power load prediction more accurately. The feature fusion network sequentially performs feature extraction in the time dimension and the channel dimension, can fuse the convolutional feature vector and the long short-term memory feature vector multi-dimensionally, and the obtained target fusion vector can comprehensively represent the short-term local features and long-term features of the historical sequence data, thereby improving the accuracy of load prediction.

[0107] The power load prediction device provided by the present invention will be described below. The power load prediction device described below can be correspondingly referred to the power load prediction method described above.

[0108] Figure 8 It is a schematic structural diagram of the power load prediction device provided by the present invention, specifically including: A data acquisition module 801, configured to acquire historical power load sequence data of a target power grid system, where the historical power load sequence data includes power load information and temperature information. For detailed description, refer to the relevant description corresponding to the above method embodiment, which will not be elaborated here.

[0109] A feature extraction module 802, configured to extract a spatial feature vector and a time feature vector of the historical power load sequence data. The spatial feature vector is a multi-channel vector obtained by performing spatial feature extraction on the historical power load sequence data, and the time feature vector is at least one-channel long short-term memory feature vector obtained by performing time feature extraction on the historical power load sequence data. For detailed description, refer to the relevant description corresponding to the above method embodiment, which will not be elaborated here.

[0110] A feature fusion module 803, configured to perform feature fusion on the spatial feature vector and the time feature vector to generate a feature fusion vector. For detailed description, refer to the relevant description corresponding to the above method embodiment, which will not be elaborated here.

[0111] A prediction module 804, configured to predict the power load information at a future time point of the target power grid system based on the feature fusion vector. For detailed description, refer to the relevant description corresponding to the above method embodiment, which will not be elaborated here.

[0112] Figure 9 Illustrates a schematic physical structure diagram of an electronic device, such as Figure 9As shown in the figure, the electronic device may include: a processor 910, a communications interface 920, a memory 930, and a communication bus 940. Among them, the processor 910, the communications interface 920, and the memory 930 communicate with each other through the communication bus 940. The processor 910 may call the logical instructions in the memory 930 to execute a power load prediction method, which includes: obtaining historical power load sequence data of a target power grid system, where the historical power load sequence data includes power load information and temperature information; extracting a spatial feature vector and a temporal feature vector of the historical power load sequence data. The spatial feature vector is a multi-channel vector obtained by performing spatial feature extraction on the historical power load sequence data, and the temporal feature vector is at least one channel of long short-term memory feature vectors obtained by performing temporal feature extraction on the historical power load sequence data; fusing the spatial feature vector and the temporal feature vector to generate a feature fusion vector; predicting the power load information at a future time point of the target power grid system based on the feature fusion vector.

[0113] In addition, when the logical instructions in the above-mentioned memory 930 are implemented in the form of software functional units and sold or used as independent products, they may be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, may be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as a USB flash drive, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk, or an optical disc that can store program codes.

[0114] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the power load prediction method provided by each of the above methods. The method includes: obtaining historical power load sequence data of a target power grid system, where the historical power load sequence data includes power load information and temperature information; extracting a spatial feature vector and a temporal feature vector from the historical power load sequence data. The spatial feature vector is a multi-channel vector obtained by performing spatial feature extraction on the historical power load sequence data, and the temporal feature vector is at least one channel of long short-term memory feature vectors obtained by performing temporal feature extraction on the historical power load sequence data; fusing the spatial feature vector and the temporal feature vector to generate a feature fusion vector; predicting the power load information at a future time point of the target power grid system based on the feature fusion vector.

[0115] On another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is configured to execute the power load prediction method provided by each of the above methods. The method includes: obtaining historical power load sequence data of a target power grid system, where the historical power load sequence data includes power load information and temperature information; extracting a spatial feature vector and a temporal feature vector from the historical power load sequence data. The spatial feature vector is a multi-channel vector obtained by performing spatial feature extraction on the historical power load sequence data, and the temporal feature vector is at least one channel of long short-term memory feature vectors obtained by performing temporal feature extraction on the historical power load sequence data; fusing the spatial feature vector and the temporal feature vector to generate a feature fusion vector; predicting the power load information at a future time point of the target power grid system based on the feature fusion vector.

[0116] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0117] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting electric power load, characterized in that, Including: Obtain historical power load sequence data of a target power grid system, where the historical power load sequence data includes power load information and temperature information; Extract the spatial feature vector and the temporal feature vector of the historical power load sequence data. The spatial feature vector is a multi-channel vector obtained by performing spatial feature extraction on the historical power load sequence data, and the temporal feature vector is at least one channel of long short-term memory feature vectors obtained by performing temporal feature extraction on the historical power load sequence data; Fuse the spatial feature vector and the temporal feature vector to generate a feature fusion vector; Predict the power load information at a future time point of the target power grid system based on the feature fusion vector.

2. The method according to claim 1, characterized in that, The obtaining of the historical power load sequence data of the target power grid system includes: Based on a preset time period, collect the historical power load data corresponding to multiple historical time points of the target power grid system to obtain historical power load sequence data.

3. The method according to claim 1, wherein The extracting of the spatial feature vector and the temporal feature vector of the historical power load sequence data includes: Perform spatial feature extraction on the historical power load sequence data through a convolutional neural network to obtain a multi-channel spatial feature vector; Perform temporal feature extraction on the historical power load sequence data through a long short-term memory network to obtain at least one channel of temporal feature vectors, and the temporal feature vectors are long short-term memory feature vectors.

4. The method according to claim 3, characterized in that The convolutional neural network includes a plurality of serially arranged convolutional units, and the convolutional unit includes a first convolutional layer, a second convolutional layer, and a third convolutional layer; The performing of spatial feature extraction on the historical power load sequence data through the convolutional neural network to obtain a multi-channel spatial feature vector includes: When there is an (i - 2)-th initial eigenvalue, perform convolutional processing on the i-th initial eigenvalue, the (i - 1)-th initial eigenvalue, and the (i - 2)-th initial eigenvalue in the initial sequence data through the first convolutional layer to generate the i-th first eigenvalue. The initial sequence data includes a plurality of initial eigenvalues, and the initial sequence data input to the first convolutional layer of the first convolutional unit is the historical power load sequence data; When there is no (i - 2)-th initial eigenvalue, perform convolutional processing on the i-th initial eigenvalue and the (i - 1)-th initial eigenvalue in the initial sequence data through the first convolutional layer to generate the i-th first eigenvalue, where i ≥ 2; Input the first sequence data including each first eigenvalue into the second convolutional layer; When there is a (j - 4)-th first eigenvalue, perform convolutional processing on the j-th first eigenvalue, the (j - 2)-th first eigenvalue, and the (j - 4)-th first eigenvalue in the first sequence data through the second convolutional layer to generate the j-th second eigenvalue; When there is no (j - 4)-th first eigenvalue, perform convolutional processing on the j-th first eigenvalue and the (j - 2)-th first eigenvalue in the first sequence data through the second convolutional layer to generate the j-th second eigenvalue, where j ≥ 4; Input the second sequence data including each second eigenvalue into the third convolutional layer; When there is a (k - 6)-th second eigenvalue, the third convolutional layer performs convolutional processing on the k-th second eigenvalue, the (k - 3)-th second eigenvalue, and the (k - 6)-th second eigenvalue in the second sequence data to generate a k-th third eigenvalue; When there is no (k - 6)-th second eigenvalue, the third convolutional layer performs convolutional processing on the k-th second eigenvalue and the (k - 3)-th second eigenvalue in the second sequence data to generate a k-th third eigenvalue, obtaining third sequence data of each third eigenvalue, where k ≥ 7; Based on the third sequence data output by the last convolutional unit, a multi-channel spatial feature vector is obtained.

5. The method according to claim 3, wherein The step of fusing the spatial feature vector and the temporal feature vector to generate a feature fusion vector includes: The feature fusion network concatenates the spatial feature vector and the temporal feature vector along the channel dimension to generate a feature fusion vector.

6. The method according to claim 1 or 5, characterized in that The step of predicting the future time point's power load information of the target power grid system based on the feature fusion vector includes: Performing feature extraction on the feature fusion vector in the time dimension to obtain an intermediate fusion vector; Performing feature extraction on the intermediate fusion vector in the channel dimension to obtain a target fusion vector; Predicting the future time point's power load information of the target power grid system according to the target fusion vector.

7. The method according to claim 6, wherein The step of performing feature extraction on the feature fusion vector in the time dimension to obtain an intermediate fusion vector and performing feature extraction on the intermediate fusion vector in the channel dimension to obtain a target fusion vector includes: Performing a transpose process on the feature fusion vector to obtain a first feature vector of C×T; Performing feature extraction on the first feature vector row by row to obtain a second feature vector of C×T; Performing a transpose process on the second feature vector to obtain a third feature vector of T×C, and generating the intermediate fusion vector according to the third feature vector; Performing feature extraction on the intermediate fusion vector row by row to generate a fourth feature vector of T×C, and generating the target fusion vector according to the fourth feature vector; The size of the feature fusion vector is T×C, and C = c1 + c2, where c1 is the number of channels of the spatial feature vector, c1 ≥ 2, c2 is the number of channels of the temporal feature vector, c2 ≥ 1, and T is the number of eigenvalues of the spatial feature vector and the temporal feature vector.

8. An electric load prediction device, characterized in that, It includes: A data acquisition module for acquiring historical power load sequence data of the target power grid system, where the historical power load sequence data includes power load information and temperature information; A feature extraction module for extracting the spatial feature vector and the temporal feature vector of the historical power load sequence data. The spatial feature vector is a multi-channel vector obtained by performing spatial feature extraction on the historical power load sequence data, and the temporal feature vector is at least one-channel long short-term memory feature vector obtained by performing temporal feature extraction on the historical power load sequence data; A feature fusion module for fusing the spatial feature vector and the temporal feature vector to generate a feature fusion vector; A prediction module, configured to predict power load information at a future time point of the target power grid system based on the feature fusion vector.

9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that, When the processor executes the computer program, the power load prediction method according to any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, the power load prediction method according to any one of claims 1 to 7 is implemented.

Images