A power load prediction method based on time series core fusion

Abstract

This invention provides a power load forecasting method based on time series core fusion, belonging to the field of artificial intelligence technology. The method involves normalizing adjustable load data to obtain adjustable load data feature vectors; constructing an adjustable power load forecasting model using a multi-layer perception layer, a random pooling layer, and a fusion layer; inputting the adjustable load data feature vectors into the adjustable power load forecasting model; training the adjustable power load forecasting model based on an accuracy index; and finally, obtaining the adjustable power load forecasting result by performing inverse normalization calculation on the feature sequences containing the forecast group, thus completing the forecasting of adjustable power load. This invention solves the problem of low accuracy in adjustable power load forecasting results.

Description

Technical Field

[0001] This specification relates to the field of artificial intelligence technology, and in particular to a power load forecasting method based on time series core fusion. Background Technology

[0002] Adjustable loads, compared to traditional electrical loads, possess dual characteristics of both power supply and user. Their load characteristics can be decomposed into two parts: a periodic component of equal magnitude and an aperiodic nonlinear component. Currently, the field of adjustable load forecasting primarily utilizes traditional load forecasting methods such as regression forecasting, smoothing forecasting, and LSTM forecasting. However, these methods lack universality in adjustable load forecasting; as the aperiodic nonlinear component increases, the prediction accuracy significantly decreases. While unit-time neural network methods, represented by LSTM forecasting, can enhance prediction accuracy through network structure enhancement and overcome the increased error caused by large aperiodic nonlinear components, they significantly increase computation time and computational resource requirements. Summary of the Invention

[0003] To address the aforementioned shortcomings in existing technologies, this invention provides a power load forecasting method based on time series core fusion, which solves the problem of low accuracy in adjustable power load forecasting results.

[0004] To achieve the aforementioned objectives, the technical solution adopted by this invention is: a power load forecasting method based on time series core fusion, comprising:

[0005] S1: By normalizing the adjustable load data, an adjustable load data feature vector is obtained; wherein, the adjustable load data feature vector is a time series feature vector including low-frequency components and high-frequency components;

[0006] S2: Construct an adjustable power load prediction model using a multi-layer sensing layer, a random pooling layer, a Kalman prediction layer, and a fusion layer;

[0007] S3: Input the adjustable load data feature vector into the adjustable power load prediction model, and train the adjustable power load prediction model based on the accuracy index to obtain a trained adjustable power load prediction model; wherein, the trained adjustable power load prediction model is used to analyze the adjustable load data feature vector to obtain a feature sequence containing prediction groups.

[0008] S4: By performing inverse normalization calculation on the feature sequence containing the prediction group, the adjustable power load prediction result is obtained, thus completing the prediction of the adjustable power load.

[0009] Furthermore, the expression for the adjustable load data feature vector is:

[0010] ;

[0011] ;

[0012] ;

[0013] in, This represents the feature vector of adjustable load data. Indicates the normalized weights. This indicates adjustable load data. This represents the average value of adjustable load data. This represents the standard deviation of adjustable load data. Indicates the normalized offset. Indicates the total duration. express The feature vector of adjustable load data at any given time.

[0014] Furthermore, the adjustable power load forecasting model includes:

[0015] A multi-layer perception layer is used to perform linear and dimensional transformations on the feature vector of the adjustable load data to obtain the transformed prediction sequence.

[0016] A random pooling layer is used to learn features of the adjustable load data based on the predicted sequence and generate a feature transition matrix;

[0017] The Kalman prediction layer is used to fuse high- and low-frequency sequences in the feature transition matrix to obtain the feature core matrix.

[0018] The fusion layer includes a feature core matrix, which is used to fuse the feature core matrix with the adjustable load data feature vector to obtain a feature sequence containing prediction groups.

[0019] Further, S3 includes:

[0020] The adjustable load data feature vector is used as the initial variable of the adjustable power load prediction model:

[0021] ;

[0022] Using the initial variables, the feature core matrix is ​​obtained through multiple perceptron layers and random pooling layers:

[0023] ;

[0024] The feature core matrix is ​​fused with the adjustable load data feature vector to obtain a feature sequence containing the prediction group:

[0025] ;

[0026] The error between the true feature sequence and the feature sequence containing the prediction group is calculated to obtain an accuracy index. This index is then used to train the adjustable power load prediction model, resulting in a trained adjustable power load prediction model.

[0027] ;

[0028] ;

[0029] in, These represent the initial variables of the adjustable power load forecasting model. This represents the feature vector of adjustable load data. Represents the feature core matrix, express Wheel feature transition matrix, and Both represent discrete time points. This represents the low-frequency component extraction function. express Wheel feature transition matrix, This represents the high-frequency component extraction function. Indicates the number of iterations. Indicates the total duration. Indicates the additional time length. This represents a feature sequence containing the prediction group. This represents the feature sequence containing the prediction group in the (i-1)th round. This represents the intermediate parameters used to calculate the feature core matrix. This represents a feature sequence containing the prediction group. express Time-of-flight feature core matrix This represents the mean square error. Indicates the number of feature sequences. Represents the true feature sequence. Indicates the mean absolute error. Indicates the iteration round.

[0030] Furthermore, the expression for the adjustable power load forecast result is as follows:

[0031] ;

[0032] in, This indicates the results of adjustable power load forecasting. This represents the standard deviation of adjustable load data. This represents a feature sequence containing the prediction group. Indicates the normalized offset. Indicates the normalized weights. This represents the average value of adjustable load data.

[0033] The beneficial effects of the present invention are as follows: the adjustable load data is normalized to obtain the adjustable load data feature vector, and the adjustable power load prediction model is used to analyze the adjustable load data feature vector to obtain the adjustable power load prediction result. (1) Based on the sequence mean and standard deviation, a learnable linear parameter is introduced at the same time, so that the entire normalization and denormalization process can be continuously iteratively optimized according to the network training process, thereby obtaining the best normalization effect, improving the network's operational stability and improving the prediction accuracy; (2) The adjustable power load prediction model is used to aggregate information from different sequence features to obtain the global core representation, and the local sequence representation is fused with the core representation to realize the indirect interaction between sequence features. This centralized interaction not only reduces the comparison complexity, but also utilizes the aggregated representation, which helps to improve the local representation and improve the network's computation speed and accuracy; (3) Using the multi-layer perception layer and random pooling layer, while ensuring the computation speed, the core features of each feature sequence are learned and fused and assigned to each feature sequence, so that the generated data has a very high correlation with the input data and can better fit the future actual value. Attached Figure Description

[0034] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein:

[0035] Figure 1 This is an exemplary flowchart of a power load forecasting method based on time series core fusion, as shown in some embodiments of this specification.

[0036] Figure 2 This is an exemplary schematic diagram of adjustable load data with two characteristic components, shown according to some embodiments of this specification;

[0037] Figure 3 This is an exemplary schematic diagram of an adjustable load data feature vector according to some embodiments of this specification;

[0038] Figure 4(a) is an exemplary schematic diagram of adjustable power load forecasting results according to some embodiments of this specification;

[0039] Figure 4(b) is an exemplary schematic diagram of adjustable power load forecasting results according to some embodiments of this specification;

[0040] Figure 5This is an exemplary schematic diagram illustrating the error between the first feature component and the second feature component according to some embodiments of this specification. Detailed Implementation

[0041] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0042] Example

[0043] Figure 1 This is an exemplary flowchart illustrating a power load forecasting method based on time series core fusion, according to some embodiments of this specification. Figure 1 As shown, the process includes the following steps. In some embodiments, the process may be executed by a processor.

[0044] S1: By normalizing the adjustable load data, the feature vector of the adjustable load data is obtained.

[0045] Adjustable load data is data on how the load of a power system changes over time.

[0046] In some embodiments, the processor can acquire adjustable load data based on historical load data recorded by the power system.

[0047] Adjustable load data feature vectors are normalized adjustable load data vector forms. For example, adjustable load data feature vectors can be time-series power load feature vectors; adjustable load data feature vectors can include low-frequency components and high-frequency components.

[0048] In some embodiments, the expression for the adjustable load data feature vector can be:

[0049] ;

[0050] ;

[0051] ;

[0052] in, This represents the feature vector of adjustable load data. Indicates the normalized weights. This indicates adjustable load data. This represents the average value of adjustable load data. This represents the standard deviation of adjustable load data. Indicates the normalized offset. Indicates the total duration. express The feature vector of adjustable load data at any given time.

[0053] In some embodiments, the processor can perform wavelet transform on the feature vector of the adjustable load data, extract feature sequences by frequency, and obtain low-frequency and high-frequency components.

[0054] The low-frequency component is the low-frequency component in the characteristic vector of adjustable load data that reflects the long-term trend.

[0055] High-frequency components are the high-frequency components in the characteristic vector of adjustable load data that reflect short-term fluctuations.

[0056] In some embodiments, the expressions for the low-frequency components and the high-frequency components can be:

[0057] ;

[0058] ;

[0059] ;

[0060] ;

[0061] in, Indicates low-frequency components. This represents the low-frequency component extraction function. and Both represent discrete time points. Indicates the total duration. Represents high-frequency components. This represents the high-frequency component extraction function.

[0062] S2: An adjustable power load prediction model is constructed using a multi-layer sensing layer, a random pooling layer, and a feature core matrix.

[0063] Adjustable power load forecasting models are used to predict power system load and obtain power load forecast results. There are various types of adjustable power load forecasting models. For example, adjustable power load forecasting models can include deep neural network models.

[0064] In some embodiments, the input to the adjustable power load forecasting model may be an adjustable load data feature vector, and the output of the adjustable power load forecasting model may be a feature sequence containing a forecast group.

[0065] In some embodiments, the structure of the adjustable power load forecasting model is as follows:

[0066] The adjustable power load forecasting model comprises a multi-layer sensing layer, a stochastic pooling layer, a Kalman prediction layer, and a fusion layer. The output of the multi-layer sensing layer serves as the input to the stochastic pooling layer, the output of the stochastic pooling layer serves as the input to the Kalman prediction layer, the outputs of the multi-layer sensing layer and the Kalman prediction layer serve as the input to the fusion layer, and the output of the fusion layer serves as the final output of the adjustable power load forecasting model.

[0067] The multilayer perceptron is used to perform linear and dimensional transformations on the feature vectors of adjustable load data to obtain the transformed prediction sequence. The input of the multilayer perceptron can include the feature vectors of adjustable load data, and the output can include the prediction sequence.

[0068] The prediction sequence is a sequence of feature vectors of adjustable load data after linear dimensionality transformation.

[0069] In some embodiments, the multilayer sensing layer may include an input module, a first fully connected module, a second fully connected module, and an output module.

[0070] In some embodiments, the input module can be used to obtain the feature vector of adjustable load data, the first fully connected module and the second fully connected module can be used to analyze the feature vector of adjustable load data to obtain the analysis result, and the output module can be used to output the analysis result.

[0071] In some embodiments, the activation function used by the first fully connected module and the second fully connected module can be the GELU activation function:

[0072] ;

[0073] in, This represents the output of the activation function. This represents the feature values ​​after passing through the network activation layer. The distribution rules representing the characteristic sequence, The cumulative distribution function represents the normal distribution.

[0074] Random pooling layers are used to learn features of adjustable load data based on predicted sequences, generating a feature kernel matrix. The input to a random pooling layer can include the predicted sequence, and the output can include a feature transition matrix.

[0075] The feature transition matrix is ​​a transition matrix that reflects the feature information of high-frequency component sequences and low-frequency component sequences.

[0076] In some embodiments, the expression for the characteristic transition matrix can be:

[0077] ;

[0078] ;

[0079] in, This represents the random pooling layer function. Represents the multi-sensor layer function. This represents the feature sequence of the prediction group labeled i-1. This represents the intermediate parameters used to calculate the feature core matrix. This represents the coefficients used to calculate the characteristic transition matrix. This represents the value of the element in row i and column j of A. Indicates the number of features. Indicates the first in A The value of the element in row j. Indicates the first OK, This represents an associative function. This represents the characteristic transition matrix.

[0080] In some embodiments, The shape can be It includes high-frequency component sequence feature information and low-frequency component sequence feature information; among which, Represents the real number field. Indicates the number of sequence features. Indicates the total duration.

[0081] The Kalman prediction layer is used to fuse high- and low-frequency sequences in the feature transition matrix to obtain the feature core matrix. The input of the Kalman prediction layer can include the feature transition matrix, and the output can include the feature core matrix.

[0082] In some embodiments, the processor can input the predicted sequence into the Kalman prediction layer and obtain the feature core matrix through inverse wavelet transform. For example, the processor can calculate the state transition matrix based on the predicted sequence; use the state transition matrix to calculate the prediction result; and obtain the feature core matrix by fusing high and low frequencies of the prediction result.

[0083] The state transition matrix is ​​a matrix that reflects the state transitions of the characteristic core matrix.

[0084] In some embodiments, the expression for the state transition matrix can be:

[0085] ;

[0086] in, Represents the state transition matrix. express The feature core matrix, express The feature core matrix, express The feature core matrix.

[0087] The prediction results reflect the transformation of the core feature matrix over a future period.

[0088] In some embodiments, the expression for the prediction result can be:

[0089] ;

[0090] in, Indicates the prediction result. Represents the state transition matrix raised to the power of t. This represents the data with index T in the feature core matrix.

[0091] The fusion layer is used to fuse the feature core matrix with the adjustable load data feature vector to obtain a feature sequence containing the prediction group. The input to the fusion layer can include the feature core matrix and the adjustable load data feature vector, and the output can include the feature sequence containing the prediction group.

[0092] The feature core matrix is ​​a matrix that reflects the feature information of the aggregated sequence.

[0093] In some embodiments, the processor can use a feature core matrix to perform core fusion on the adjustable load data feature vectors to obtain a feature sequence containing prediction groups. For example, the processor can assign the Nth row of the feature core matrix to the Mth adjustable load data feature vector, and the combination method is to directly add the assigned sequence to the corresponding normalized feature sequence to obtain a feature sequence containing prediction groups.

[0094] S3: Input the adjustable load data feature vector into the adjustable power load prediction model, and train the adjustable power load prediction model based on the accuracy index to obtain a trained adjustable power load prediction model; wherein, the trained adjustable power load prediction model is used to analyze the adjustable load data feature vector to obtain a feature sequence containing prediction groups.

[0095] In some embodiments, the adjustable power load forecasting model can be trained using multiple labeled training samples. For example, multiple labeled training samples can be input into an initial adjustable power load forecasting model. An accuracy index is constructed using the labels and the results of the initial adjustable power load forecasting model. Based on the accuracy index, the parameters of the initial adjustable power load forecasting model are iteratively updated using gradient descent or other methods. When preset conditions are met, model training is complete, resulting in a trained adjustable power load forecasting model. These preset conditions may include accuracy index convergence, the number of iterations reaching a threshold, etc.

[0096] In some embodiments, training samples may include feature vectors of historical adjustable load data. Labels may be corresponding ground truth feature sequences. Labels may be manually labeled.

[0097] The true feature sequence is a sequence constructed from the actual monitored adjustable power load results at corresponding time points in the power system.

[0098] In some embodiments, the processor may implement S3 based on the following steps: using the adjustable load data feature vector as the initial variable of the adjustable power load prediction model; using the initial variable, obtaining a feature core matrix through a multi-layer sensing layer and a random pooling layer; fusing the feature core matrix with the adjustable load data feature vector to obtain a feature sequence containing prediction groups; calculating the error between the true feature sequence and the feature sequence containing prediction groups to obtain an accuracy index; training the adjustable power load prediction model to obtain a trained adjustable power load prediction model.

[0099] The initial variables are the input variables of the adjustable power load forecasting model before training.

[0100] In some embodiments, the expression for the initial variable can be:

[0101] .

[0102] In some embodiments, the expression for the feature core matrix can be:

[0103] .

[0104] In some embodiments, the expression containing the feature sequence of the prediction group can be:

[0105] .

[0106] Accuracy metrics are indicators used to evaluate the performance of adjustable power load forecasting models. For example, accuracy metrics may include mean squared error and mean absolute error.

[0107] In some embodiments, the expression for the accuracy metric can be:

[0108] ;

[0109] ;

[0110] in, These represent the initial variables of the adjustable power load forecasting model. This represents the feature vector of adjustable load data. Represents the feature core matrix, express Wheel feature transition matrix, and Both represent discrete time points. This represents the low-frequency component extraction function. express Wheel feature transition matrix, This represents the high-frequency component extraction function. Indicates the number of iterations. Indicates the total duration. Indicates the additional time length. This represents a feature sequence containing the prediction group. This represents the feature sequence containing the prediction group in the (i-1)th round. This represents the intermediate parameters used to calculate the feature core matrix. This represents a feature sequence containing the prediction group. express Time-of-flight feature core matrix This represents the mean square error. Indicates the number of feature sequences. Represents the true feature sequence. Indicates the mean absolute error. Indicates the iteration round; where, .

[0111] S4: By performing inverse normalization calculation on the feature sequence containing the prediction group, the adjustable power load prediction result is obtained, thus completing the prediction of the adjustable power load.

[0112] Adjustable power load forecast results reflect the adjustable power load of the power system over a future period of time.

[0113] In some embodiments, the expression for the adjustable power load forecast result can be:

[0114] ;

[0115] in, This indicates the results of adjustable power load forecasting. This represents the standard deviation of adjustable load data. This represents a feature sequence containing the prediction group. Indicates the normalized offset. Indicates the normalized weights. This represents the average value of adjustable load data.

[0116] In some embodiments, the processor can use the first 90 time-series points of adjustable load data with two characteristic components as prediction data points, and normalize the data to obtain the following: Figure 3The raw data used for prediction (i.e., the adjustable load data feature vector) shown is processed by a multilayer perceptron and a random pooling layer to generate a feature core matrix. Finally, the feature core matrix is ​​assigned to each feature sequence and then denormalized to obtain the adjustable power load prediction results shown in Figures 4(a) and 4(b); where, as Figure 2 As shown, adjustable load data with two characteristic components can include a first characteristic component (periodic load), a second characteristic component (aperiodic oscillation load), a total load time series length (100 units), and 1 hour for each unit.

[0117] In some embodiments, such as Figure 5 As shown, the sum of the errors of the first and second feature components has an average error of 0.0015 in the distribution across time series 91 to 100, with an accuracy of 99.25%. The accuracy comparison results of using other methods to predict the data used in this method are shown in Table 1.

[0118] Table 1

[0119]

[0120] In some embodiments of this specification, the processor can normalize the adjustable load data to obtain the adjustable load data feature vector, and use the adjustable power load prediction model to analyze the adjustable load data feature vector to obtain the adjustable power load prediction result. (1) Based on the sequence mean and standard deviation, a learnable linear parameter is introduced to enable the entire normalization and denormalization process to be continuously iteratively optimized according to the network training process, thereby obtaining the best normalization effect, improving the network's operational stability, and improving the prediction accuracy; (2) The adjustable power load prediction model aggregates information from different sequence features to obtain the global core representation, and merges the local sequence representation with the core representation to realize the indirect interaction between sequence features. This centralized interaction not only reduces the comparison complexity, but also utilizes the aggregated representation, which helps to improve the local representation and improve the network's computation speed and accuracy; (3) Using a multi-layer perceptual layer and a random pooling layer, while ensuring computation speed, the core features of each feature sequence are learned and then merged and assigned to each feature sequence, so that the generated data has a very high correlation with the input data and can better fit the future actual value.

Claims

1. A method for adjustable power load forecasting based on time series core fusion, characterized in that, include: S1: By normalizing the adjustable load data, an adjustable load data feature vector is obtained; wherein, the adjustable load data feature vector is a time series feature vector including low-frequency components and high-frequency components; S2: Construct an adjustable power load prediction model using a multi-layer sensing layer, a random pooling layer, a Kalman prediction layer, and a fusion layer; The adjustable power load prediction model includes: A multi-layer perception layer is used to perform linear and dimensional transformations on the feature vector of the adjustable load data to obtain the transformed prediction sequence. A random pooling layer is used to learn features of adjustable load data based on the predicted sequence and generate a feature transition matrix. The Kalman prediction layer is used to fuse high- and low-frequency sequences in the feature transition matrix to obtain the feature core matrix. The fusion layer includes a feature core matrix, which is used to fuse the feature core matrix with the adjustable load data feature vector to obtain a feature sequence containing prediction groups; S3: Input the adjustable load data feature vector into the adjustable power load prediction model, and train the adjustable power load prediction model based on the accuracy index to obtain a trained adjustable power load prediction model; wherein, the trained adjustable power load prediction model is used to analyze the adjustable load data feature vector to obtain a feature sequence containing prediction groups; including: The adjustable load data feature vector is used as the initial variable of the adjustable power load prediction model: ； Using the initial variables, the feature core matrix is ​​obtained through multiple perceptron layers and random pooling layers: ； The feature core matrix is ​​fused with the adjustable load data feature vector to obtain a feature sequence containing the prediction group: ； The error between the true feature sequence and the feature sequence containing the prediction group is calculated to obtain an accuracy index. This index is then used to train the adjustable power load prediction model, resulting in a trained adjustable power load prediction model. ； ； in, These represent the initial variables of the adjustable power load forecasting model. This represents the feature vector of adjustable load data. Represents the feature core matrix, express Wheel feature transition matrix, and Both represent discrete time points. This represents the low-frequency component extraction function. express Wheel feature transition matrix, This represents the high-frequency component extraction function. Indicates the number of iterations. Indicates the total duration. Indicates the additional time length. This represents a feature sequence containing the prediction group. This represents the feature sequence containing the prediction group in the (i-1)th round. These represent the intermediate parameters used to calculate the core feature matrix. This represents a feature sequence containing the prediction group. express Time-of-flight feature core matrix Indicates mean square error. Indicates the number of feature sequences. Represents the true feature sequence. Indicates the mean absolute error. Indicates the iteration round; S4: By performing inverse normalization calculation on the feature sequence containing the prediction group, the adjustable power load prediction result is obtained, thus completing the prediction of the adjustable power load.

2. The power load forecasting method based on time series core fusion according to claim 1, characterized in that, The expression for the adjustable load data feature vector is: ； ； ； in, This represents the feature vector of adjustable load data. Indicates the normalized weights. This indicates adjustable load data. This represents the average value of adjustable load data. This represents the standard deviation of adjustable load data. Indicates the normalized offset. Indicates the total duration. express The feature vector of adjustable load data at any given time. 3.The power load forecasting method based on time series core fusion according to claim 1, characterized in that, The expression for the adjustable power load forecast result is as follows: ； in, This indicates the results of adjustable power load forecasting. This represents the standard deviation of adjustable load data. This represents a feature sequence containing the prediction group. Indicates the normalized offset. Indicates the normalized weights. This represents the average value of adjustable load data.

Images