Short-term power load prediction method based on hybrid dilated convolutional neural network

Abstract

The invention relates to the technical field of power load prediction, in particular to a short-term power load prediction method based on a hybrid cavity convolutional neural network, which comprises the following steps: acquiring original power load data, and constructing a load sequence data set; converting a load sequence in the load sequence data set into a Grubm angle field to obtain preprocessed load input data; constructing a load prediction model based on the hybrid cavity convolutional neural network, and training the load prediction model by using the preprocessed load input data; and collecting power load data in real time, and performing power load prediction by using the trained load prediction model. According to the method, the load sequence is converted into the Grubrum angle field to extract the fine-grained features of the load sequence, the fine-grained features are used as the input of the load prediction model, complex time sequence features can be captured, abnormal changes in the load sequence can be identified, the fine changes of the load in a short time can be effectively reflected, and the accuracy of short-term power load prediction can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power load forecasting, and in particular to a short-term power load forecasting method based on a hybrid atrous convolutional neural network. Background Art

[0002] Power load forecasting technology is based on various historical data, such as temperature, humidity, historical load, etc., to predict the load data at a certain time in the future. According to the forecast time scale, power load forecasting can be divided into long-term forecasting, medium-term forecasting and short-term forecasting. Short-term load forecasting refers to daily load forecasting and weekly load forecasting, which is the key to demand-side power management. Its purpose is to promote the safe and stable operation of modern power systems. Accurate short-term forecasting provides an important basis for power dispatching, power supply planning, safety monitoring, and unit optimization, which is conducive to the safer and more economical operation of the power system and improves the economic and social benefits of power grid operation.

[0003] At present, short-term power load forecasting methods are mainly divided into two categories. One is the regression analysis method based on time series and the fuzzy prediction method. These methods have simple calculation principles and fast prediction speed, but they have high requirements for historical data, low prediction accuracy, and cannot reflect the periodic correlation of time series. The other method is the intelligent prediction method represented by neural network, and its application is mainly divided into convolutional neural network and recurrent neural network. This method uses the power load of the past period as training samples, constructs a suitable network structure, and outputs the load value at a certain moment in the future after meeting the accuracy requirements through multiple rounds of training. The prediction method based on neural network can have good adaptive ability for a large number of non-structural and non-precise load sequences, and the prediction accuracy is significantly improved compared with the first method. However, its shortcomings are that the number of parameters involved in the prediction model is large and the calculation overhead is large. Among them, recurrent neural network can only be used for serial calculation, and cannot process load sequences in parallel, which is not conducive to the rapid convergence of the model. The traditional convolutional neural network is limited by computing resources and cannot further increase the number and size of convolution kernels, resulting in difficulty in improving model performance. In addition, the traditional load forecasting method directly takes the one-dimensional sequence of load as input. Although it can intuitively reflect the fluctuation of power load, it cannot fully reflect the fine-grained load characteristics, which is not conducive to improving the accuracy of the prediction model. Summary of the invention

[0004] To this end, the technical problem to be solved by the present invention is to overcome the problem that the prior art cannot fully reflect fine-grained load characteristics and has high computational overhead, resulting in insufficient accuracy in short-term power load forecasting.

[0005] In order to solve the above technical problems, the present invention provides a short-term power load forecasting method based on a hybrid atrous convolutional neural network, comprising:

[0006] Collect original power load data and construct load sequence data set;

[0007] Convert the load sequence in the load sequence data set into a Gram angle field to obtain preprocessed load input data;

[0008] A load forecasting model is constructed based on a hybrid atrous convolutional neural network, and the load forecasting model is trained using preprocessed load input data.

[0009] Collect power load data in real time and use the trained load forecasting model to forecast power load.

[0010] Preferably, the process of constructing a load sequence data set includes:

[0011] Perform data cleaning on the original power load data to obtain a load sequence data set;

[0012] The data cleaning of the original power load data includes: checking data consistency and removing invalid values ​​and missing values.

[0013] Preferably, converting the load sequence in the load sequence data set into a Gram angle field to obtain preprocessed load input data includes:

[0014] Normalize the load series in the load series data set;

[0015] The normalized load sequence is converted from the Cartesian coordinate system to the polar coordinate system to obtain the angular cosine corresponding to each load value in the normalized load sequence;

[0016] According to the angular cosine corresponding to each load value in the normalized load sequence, the Gram angle sum field matrix and the Gram angle difference field matrix of the load sequence are obtained, and the Gram angle sum field matrix and the Gram angle difference field matrix are converted into two-dimensional images as preprocessed load input data.

[0017] Preferably, the normalizing the load sequence in the load sequence data set includes:

[0018] The load sequence is normalized using the maximum and minimum normalization formula to map the load value to the [-1,1] amplitude range. The formula is:

[0019]

[0020] Where V represents the load sequence, v i is the load value at the i-th moment in the load sequence, i = 1, 2, ..., n, n represents the length of the load sequence, v' i v i The normalized value of .

[0021] Preferably, the normalized load sequence is converted from a Cartesian coordinate system to a polar coordinate system using the formula:

[0022]

[0023] Where V' represents the normalized load sequence, represents the cosine of the angle corresponding to the load value at the i-th moment; t i represents the time axis of the i-th moment in the load sequence, r i It represents the radius of the time axis corresponding to the i-th moment in the load sequence, and N represents the constant factor.

[0024] Preferably, according to the angular cosines corresponding to each load value in the normalized load sequence, the Gram angle sum field matrix and the Gram angle difference field matrix of the load sequence are obtained, and the formula includes:

[0025]

[0026] Where GASF represents the Gram angle and field matrix, s ij represents the element in the i-th row and j-th column of the Gram angle and field matrix, GADF represents the Gram angle difference field matrix, d ij represents the element in the i-th row and j-th column of the Gram angle difference field matrix, and They represent the angular cosines of the load values ​​at the i-th moment and the j-th moment respectively, i, j = 1, 2, …, n, where n represents the length of the load sequence.

[0027] Preferably, the load forecasting model comprises an input layer, a convolutional pooling module, a fully connected module and an output layer; the convolutional pooling module comprises a parallel Gram angle and field mixed hole convolution module and a Gram angle difference field mixed hole convolution module;

[0028] The Gram angle sum field matrix and the Gram angle difference field matrix are converted into two-dimensional images and then input into the load forecasting model, and then respectively input into the Gram angle sum field mixed hole convolution module and the Gram angle difference field mixed hole convolution module to obtain the Gram angle sum field output feature and the Gram angle difference field output feature;

[0029] The Gram angle sum field output characteristics and the Gram angle difference field output characteristics are input into the fully connected module for activation calculation to obtain the load prediction value.

[0030] Preferably, the Gram angle and field mixed dilated convolution module and the Gram angle difference field mixed dilated convolution module each include three mixed dilated convolution submodules connected in sequence, and each mixed dilated convolution submodule includes a mixed dilated convolution and a pooling layer connected in sequence;

[0031] The sampling frequencies of the mixed dilated convolution in the Gram angle and field mixed dilated convolution module are 1, 2, and 5 respectively; the sampling frequencies of the mixed dilated convolution in the Gram angle difference field mixed dilated convolution module are 1, 2, and 5 respectively.

[0032] Preferably, the pooling layer in the hybrid atrous convolution submodule adopts global maximum pooling.

[0033] Preferably, the fully connected module includes two fully connected layers, each of which includes a fully connected unit, a normalization unit, and an activation unit Dropout function that are connected in sequence.

[0034] The above technical solution of the present invention has the following beneficial effects compared with the prior art:

[0035] The short-term power load forecasting method based on a hybrid atrous convolutional neural network described in the present invention converts the load sequence into a Gram angle field to extract the fine-grained features of the load sequence as the input of the load forecasting model, which can capture complex time series features, identify abnormal changes in the load sequence, effectively reflect the subtle changes in the load in a short period of time, and improve the accuracy of short-term power load forecasting. At the same time, the present invention uses a hybrid atrous convolutional neural network to construct a load forecasting model and perform load forecasting. Compared with the traditional convolutional neural network model, the hybrid atrous convolutional neural network increases the receptive field of the convolution kernel, reduces the training time, alleviates the gradient vanishing and gradient explosion problems to a certain extent, saves computing resources, and can further improve the prediction accuracy and efficiency of load forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:

[0037] Figure 1 It is a flow chart of a short-term power load forecasting method based on a hybrid atrous convolutional neural network of the present invention;

[0038] Figure 2 is a flow chart for converting a load sequence in a load sequence data set into a Gram angle field;

[0039] Figure 3 It is a Gram angle field example diagram of a preferred embodiment of the present invention;

[0040] Figure 4 It is a structural diagram of the load forecasting model of the present invention. DETAILED DESCRIPTION

[0041] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0042] Reference Figure 1 As shown, the present invention provides a short-term power load forecasting method based on a hybrid atrous convolutional neural network, the steps comprising:

[0043] S1. Collect original power load data, clean the original power load data, remove redundant information, and obtain a load sequence data set;

[0044] S2, converting the load sequence in the load sequence data set into a Gram angle field to obtain preprocessed load input data;

[0045] S3, constructing a load forecasting model based on a hybrid atrous convolutional neural network, and using the preprocessed load input data to train the load forecasting model;

[0046] S4. Collect power load data in real time and use the trained load prediction model to perform power load prediction.

[0047] Specifically, in S1, the original power load data is cleaned, including checking data consistency, removing invalid values ​​and missing values, etc. Data cleaning can improve the quality of load data and reduce the probability of errors in load data processing.

[0048] Specifically, refer to Figure 2 , S2 includes the following steps:

[0049] S2-1. Normalize the load sequence in the load sequence data set, including:

[0050] The load sequence is normalized using the maximum and minimum value normalization formula to map the load value to the [-1,1] amplitude range. The formula is:

[0051]

[0052] Where V represents the load sequence, v i is the load value at the i-th moment in the load sequence, i = 1, 2, ..., n, n represents the length of the load sequence, v' i v i The normalized value of .

[0053] S2-2. Convert the normalized load sequence to the polar coordinate system, including:

[0054] Use the coordinate transformation formula to transform the normalized Cartesian coordinate system load sequence into the polar coordinate system load sequence. The formula is:

[0055]

[0056] Where V' represents the normalized load sequence, represents the cosine of the angle corresponding to the load value at the i-th moment; t i represents the time axis of the i-th moment in the load sequence, r i It represents the radius of the time axis corresponding to the i-th moment in the load sequence, and N represents a constant factor used to standardize the scale span of the polar coordinate system.

[0057] S2-3. Use the trigonometric function transformation of angle sum / difference to convert the load sequence into the Gram angle field as the pre-processed load input data.

[0058] The Gram angle field is divided into a Gram angle sum field and a Gram angle difference field, the difference between the two lies in the sine and cosine functions used when generating fine-grained features and the summation and difference relationship of phase parameters.

[0059] According to the angular cosine corresponding to each load value in the normalized load sequence, the Gram angle sum field matrix and the Gram angle difference field matrix of the load sequence are obtained. The conversion formula of the Gram angle sum field matrix and the Gram angle difference field matrix includes:

[0060]

[0061] Where GASF represents the Gram angle and field matrix, s ij represents the element in the i-th row and j-th column of the Gram angle and field matrix, GADF represents the Gram angle difference field matrix, d ij represents the element in the i-th row and j-th column of the Gram angle difference field matrix, and They represent the angular cosines of the load values ​​at the i-th moment and the j-th moment respectively, i, j = 1, 2, …, n, where n represents the length of the load sequence.

[0062] Converting the Gram angle sum field matrix and the Gram angle difference field matrix into images can realize the transformation of one-dimensional load sequence to two-dimensional image. The Gram angle sum field matrix reflects the overall change trend of the load sequence, and the Gram angle difference field matrix reflects the local differences and changes of the load sequence. The combination of the two is the fine-grained feature of the load sequence. Fine-grained features can reflect subtle changes in load in a short period of time, such as peak and valley periods within the day, sudden increase in demand, etc. This is particularly helpful for short-term load forecasting, which can make the load forecasting model more sensitive to changes in load in a short period of time and reduce errors.

[0063] The present invention uses the two-dimensional image obtained by transforming the Gram angle sum field matrix and the Gram angle difference field matrix as pre-processed load input data.

[0064] Figure 3 This is an example diagram of the Gram angle field of a preferred embodiment of the present invention.

[0065] Figure 4 This is a structural diagram of the load forecasting model of the present invention, which includes an input layer, a convolutional pooling module, a fully connected module and an output layer.

[0066] The convolution pooling module includes a parallel Gram angle and field mixed dilated convolution module and a Gram angle difference field mixed dilated convolution module. The Gram angle and field mixed dilated convolution module and the Gram angle difference field mixed dilated convolution module each include three sequentially connected mixed dilated convolution submodules, and each mixed dilated convolution submodule includes a sequentially connected mixed dilated convolution and a pooling layer.

[0067] The dilated convolution is a convolution that samples the image at a sampling frequency of rate, that is, it samples once for every rate pixel in the image. When rate = 0, it is a traditional convolution, and no information is lost; when rate > 0, rate-1 zeros are inserted into the adjacent points of the traditional convolution kernel, and then the image is convolved. For example, a 3*3 convolution kernel with a hole of rate = 1 inserted has a receptive field equivalent to that of a 7*7 convolution kernel, and the number of parameters has not increased. The mixed dilated convolution uses a set of dilated convolution kernels with a common divisor of rate not greater than 1 to stack, in order to overcome the defect of dilated convolution that loses information.

[0068] Since the common divisor of the rate of the mixed void convolution is not greater than 1, the present invention selects a mixed void convolution structure with rates of 1, 2, and 5. This structure can adaptively meet the feature acquisition of the Gram angle field of the load sequence and ensure that the coherence of the receptive field of the convolution kernel is not destroyed.

[0069] Specifically, the Gram angle and field mixed dilated convolution module includes a first mixed dilated convolution submodule, a second mixed dilated convolution submodule and a third mixed dilated convolution submodule connected in sequence, the rate of the mixed dilated convolution in the first mixed dilated convolution submodule is 1, the rate of the mixed dilated convolution in the first mixed dilated convolution submodule is 2, and the rate of the mixed dilated convolution in the first mixed dilated convolution submodule is 5.

[0070] The Gram angle difference field mixed void convolution module includes a fourth mixed void convolution submodule, a fifth mixed void convolution submodule and a sixth mixed void convolution submodule connected in sequence, the rate of the mixed void convolution in the fourth mixed void convolution submodule is 1, the rate of the mixed void convolution in the fifth mixed void convolution submodule is 2, and the rate of the mixed void convolution in the sixth mixed void convolution submodule is 5.

[0071] In the hybrid atrous convolution submodule, the pooling layer uses global maximum pooling to take the maximum value of feature points in adjacent areas and retain local features to reduce the dimensionality of the output features of the hybrid atrous convolution.

[0072] The fully connected module includes two fully connected layers, each of which includes a fully connected unit, a normalization unit, and an activation unit Dropout function that are connected in sequence.

[0073] The load forecasting model is trained using the preprocessed load input data, including:

[0074] Converting the Gram angle sum field matrix and the Gram angle difference field matrix into two-dimensional images as pre-processed load input data, wherein the two-dimensional image of the Gram angle sum field matrix is ​​used to reflect the overall change characteristics of the load sequence, and the two-dimensional image of the Gram angle difference field matrix is ​​used to reflect the local difference characteristics of the load sequence; batch inputting the pre-processed load input data into the load forecasting model;

[0075] The two-dimensional images of the Gram angle sum field matrix and the Gram angle difference field matrix are input into the Gram angle sum field mixed atrous convolution module and the Gram angle difference field mixed atrous convolution module respectively to obtain the Gram angle sum field output features and the Gram angle difference field output features;

[0076] The Gram angle sum field output characteristics and the Gram angle difference field output characteristics are input into the fully connected module for activation calculation to obtain the load prediction value.

[0077] Preferably, the load forecasting model uses the Adam optimization algorithm to accelerate model convergence during the training process.

[0078] In summary, the short-term power load forecasting method based on a hybrid atrous convolutional neural network described in the present invention converts the load sequence into a Gram angle field to extract the fine-grained features of the load sequence as the input of the load forecasting model, which can capture complex time series features, identify abnormal changes in the load sequence, effectively reflect the subtle changes in the load in a short period of time, and improve the accuracy of short-term power load forecasting. At the same time, the present invention uses a hybrid atrous convolutional neural network to construct a load forecasting model and perform load forecasting. Compared with the traditional convolutional neural network model, the hybrid atrous convolutional neural network increases the receptive field of the convolution kernel, reduces the training time, alleviates the gradient vanishing and gradient explosion problems to a certain extent, saves computing resources, and can further improve the prediction accuracy and efficiency of load forecasting.

[0079] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0080] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0081] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0082] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0083] Obviously, the above embodiments are merely examples for the purpose of clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A short-term power load forecasting method based on a hybrid atrous convolutional neural network, characterized in that: include: Collect original power load data and construct load sequence data set; Convert the load sequence in the load sequence data set into a Gram angle field to obtain preprocessed load input data; A load forecasting model is constructed based on a hybrid atrous convolutional neural network, and the load forecasting model is trained using preprocessed load input data. Collect power load data in real time and use the trained load forecasting model to forecast power load.

2. According to claim 1, a short-term power load forecasting method based on a hybrid atrous convolutional neural network is characterized in that: The process of constructing a load sequence dataset includes: Perform data cleaning on the original power load data to obtain a load sequence data set; The data cleaning of the original power load data includes: checking data consistency and removing invalid values ​​and missing values.

3. The short-term power load forecasting method based on hybrid atrous convolutional neural network according to claim 1 is characterized in that: The step of converting the load sequence in the load sequence data set into a Gram angle field to obtain preprocessed load input data includes: Normalize the load series in the load series data set; The normalized load sequence is converted from the Cartesian coordinate system to the polar coordinate system to obtain the angular cosine corresponding to each load value in the normalized load sequence; According to the angular cosine corresponding to each load value in the normalized load sequence, the Gram angle sum field matrix and the Gram angle difference field matrix of the load sequence are obtained, and the Gram angle sum field matrix and the Gram angle difference field matrix are converted into two-dimensional images as preprocessed load input data.

4. A short-term power load forecasting method based on a hybrid atrous convolutional neural network according to claim 3, characterized in that: The normalizing the load sequence in the load sequence data set includes: The load sequence is normalized using the maximum and minimum value normalization formula to map the load value to the [-1,1] amplitude range. The formula is: Where V represents the load sequence, v i is the load value at the i-th moment in the load sequence, i = 1, 2, ..., n, n represents the length of the load sequence, v' i v i The normalized value of .

5. A short-term power load forecasting method based on a hybrid atrous convolutional neural network according to claim 4, characterized in that: The normalized load sequence is converted from the Cartesian coordinate system to the polar coordinate system, and the formula is: Where V' represents the normalized load sequence, represents the cosine of the angle corresponding to the load value at the i-th moment; t i represents the time axis at the i-th moment in the load sequence, r i It represents the radius of the time axis corresponding to the i-th moment in the load sequence, and N represents the constant factor.

6. A short-term power load forecasting method based on a hybrid atrous convolutional neural network according to claim 5, characterized in that: According to the angular cosine corresponding to each load value in the normalized load sequence, the Gram angle sum field matrix and Gram angle difference field matrix of the load sequence are obtained. The formulas include: Where GASF represents the Gram angle and field matrix, s ij represents the element in the i-th row and j-th column of the Gram angle and field matrix, GADF represents the Gram angle difference field matrix, d ij represents the element in the i-th row and j-th column of the Gram angle difference field matrix, and They represent the angular cosines of the load values ​​at the i-th moment and the j-th moment respectively, i, j = 1, 2, …, n, where n represents the length of the load sequence.

7. A short-term power load forecasting method based on a hybrid atrous convolutional neural network according to claim 3, characterized in that: The load forecasting model includes an input layer, a convolutional pooling module, a fully connected module and an output layer; the convolutional pooling module includes a parallel Gram angle and field mixed hole convolution module and a Gram angle difference field mixed hole convolution module; The Gram angle sum field matrix and the Gram angle difference field matrix are converted into two-dimensional images and then input into the load forecasting model, and then respectively input into the Gram angle sum field mixed hole convolution module and the Gram angle difference field mixed hole convolution module to obtain the Gram angle sum field output feature and the Gram angle difference field output feature; The Gram angle sum field output characteristics and the Gram angle difference field output characteristics are input into the fully connected module for activation calculation to obtain the load prediction value.

8. A short-term power load forecasting method based on a hybrid atrous convolutional neural network according to claim 7, characterized in that: The Gram angle sum field mixed dilated convolution module and the Gram angle difference field mixed dilated convolution module each include three mixed dilated convolution submodules connected in sequence, and each mixed dilated convolution submodule includes a mixed dilated convolution and a pooling layer connected in sequence; The sampling frequencies of the mixed dilated convolution in the Gram angle and field mixed dilated convolution module are 1, 2, and 5 respectively; the sampling frequencies of the mixed dilated convolution in the Gram angle difference field mixed dilated convolution module are 1, 2, and 5 respectively.

9. The short-term power load forecasting method based on hybrid atrous convolutional neural network according to claim 7 is characterized in that: The pooling layer in the hybrid hole convolution submodule adopts global maximum pooling.

10. The short-term power load forecasting method based on hybrid atrous convolutional neural network according to claim 7 is characterized in that: The fully connected module includes two fully connected layers, each of which includes a fully connected unit, a normalization unit, and an activation unit Dropout function that are connected in sequence.

Images