A multi-time scale load forecasting method based on a hybrid neural network

Abstract

The application discloses a kind of multi-time scale load prediction methods based on hybrid neural network, obtain, pre-process the original load data of transformer, and establish pre-training model based on bidirectional long short-term memory network;Using fully connected neural network, the output vector of pre-training model is reduced dimension to obtain the training model after reduction dimension;Select similar historical meteorological data with the day to be predicted, and the output result of fully connected neural network is spliced after being converted into two-dimensional matrix;Convolutional neural network is used to extract the feature vector of two-dimensional matrix feature;To the training model after reduction dimension, the feature vector is input into the training model after setting, training, so that bidirectional long short-term memory network can realize day-ahead load prediction;Based on fully connected network, bidirectional long short-term memory network is connected with gate recurrent unit, and the abstract comprehensive data output by bidirectional long short-term memory network is input into gate recurrent unit, to realize week-ahead load prediction.

Description

Technical Field

[0001] This invention relates to the field of new energy power system technology, and in particular to a multi-timescale load forecasting method based on hybrid neural networks. Background Technology

[0002] In recent years, artificial intelligence (AI) technology has been applied to various fields and achieved remarkable results. Among numerous machine learning algorithms, deep learning models stand out, with their nonlinear fitting capabilities far superior to traditional prediction models. The development of communication technologies such as 5G has accelerated the digitalization of the energy industry. In areas such as load forecasting, power system data visualization, renewable energy forecasting, and where traditional methods still need improvement, new AI technologies have been widely applied. Accurate load forecasting is of great significance for maintaining the economical and reliable operation of the power system.

[0003] Load forecasting at different time scales plays a crucial supporting role in power system planning, dispatching, and the design of electricity market bidding systems. However, most existing research focuses on a single time scale, and a single model is insufficient to meet the needs of practical engineering.

[0004] Therefore, a multi-timescale load forecasting method based on hybrid neural networks is needed. Summary of the Invention

[0005] This invention provides a multi-timescale load forecasting method based on a hybrid neural network, which at least solves the technical problem in related technologies where load forecasting is carried out around a single time scale and a single model is difficult to meet the actual engineering needs.

[0006] According to one aspect of the present invention, a multi-time-scale load forecasting method based on a hybrid neural network is provided, comprising:

[0007] The original load data of the transformer is obtained, the original load data is preprocessed to obtain load data, and a pre-trained model is established based on a bidirectional long short-term memory network.

[0008] The output vector of the pre-trained model is reduced in dimensionality using a fully connected neural network to obtain the dimensionality-reduced training model.

[0009] Select historical meteorological data similar to the date to be predicted, and then concatenate the hourly meteorological data from the historical meteorological data similar to the date to be predicted with the output of the fully connected neural network to transform it into a two-dimensional matrix.

[0010] A convolutional neural network is used to extract features from the two-dimensional matrix and organize them to obtain feature vectors. The dimensionality-reduced training model is set according to the characteristics of hourly meteorological data, and the feature vectors are input into the set training model for training, so that the bidirectional long short-term memory network can achieve day-ahead load forecasting.

[0011] By connecting a bidirectional long short-term memory network to a gated loop unit using a fully connected network, the abstract and comprehensive data output from the bidirectional long short-term memory network is input into the gated loop unit to achieve week-ahead load forecasting.

[0012] Optionally, the obtained transformer load data is transformer load data at t-minute intervals within a year T.

[0013] Alternatively, a fully connected neural network can be used to reduce the dimension of the output vector of the pre-trained model to the load data at 1-hour intervals.

[0014] Optionally, concatenating hourly meteorological data with the output of a fully connected neural network to transform it into a two-dimensional matrix includes:

[0015] The feature points of the hourly meteorological data are used to form a daily feature vector α(i). The feature points of the hourly meteorological data include: the load L(i) at the hour, the highest temperature T(i) in one hour, and the highest air humidity H(i) in one hour.

[0016] The daily feature vectors α(i) are used to form the daily feature matrix A in a sliding window manner;

[0017] Transform the daily feature matrix A into a two-dimensional matrix:

[0018]

[0019] In the above formula, α(i)=(L(i), T(i), H(i)).

[0020] Optionally, the training model after dimensionality reduction can be configured according to the characteristics of hourly meteorological data, including setting the label value of the bidirectional long short-term memory network in the pre-trained model to the next day's load value, and setting the loss function to consider the error of the point prediction results and the influence of the definite integral of the load curve.

[0021] Optionally, a bidirectional long short-term memory network is connected to a gated recurrent unit based on a fully connected network, enabling the abstract and comprehensive data output by the bidirectional long short-term memory network to achieve week-ahead load forecasting, including:

[0022] Weekly meteorological data close to the week to be predicted is selected as the meteorological data for the week to be predicted. Daily timescale meteorological factors are connected to gated recurrent units through fully connected layers, that is, the abstract comprehensive data output by the bidirectional long short-term memory network is directly input into the gated recurrent unit.

[0023] The gate circulation unit is configured based on the characteristics of the meteorological data for the week to be predicted;

[0024] The gate loop unit is configured to output the load value η for day t of the next week. t :

[0025] η t =F GRU (ε t D t ,η t-1 )

[0026] In the above formula, FGRU It is a gate loop unit; Dt This is daily-scale data including weekday, weekday, holiday, daily high temperature, daily low temperature, daily high humidity, and daily low humidity on day t; η t-1 The load value at the previous moment; ε t This represents the abstract, comprehensive data processed by the bidirectional long short-term memory network on day t.

[0027] The load value for the following week is obtained through rolling prediction using the gate loop unit.

[0028] Optionally, ε t The expression is:

[0029]

[0030] In the formula, n t (i) represents the comprehensive abstract feature vector of meteorological factors at the i-th hour of day t after feature extraction via a convolutional neural network; m t (i) represents the load value for the i-th hour on day t.

[0031] According to another aspect of the present invention, a multi-time-scale load forecasting system based on a hybrid neural network is also provided, comprising:

[0032] The pre-trained model module is used to acquire the original load data of the transformer, preprocess the original load data to obtain load data, and establish a pre-trained model based on a bidirectional long short-term memory network.

[0033] The dimension reduction module is used to reduce the dimension of the output vector of the pre-trained model using a fully connected neural network to obtain the dimension-reduced training model.

[0034] The two-dimensional matrix module is used to select historical meteorological data similar to the date to be predicted, and to concatenate the hourly meteorological data in the historical meteorological data similar to the date to be predicted with the output of the fully connected neural network to transform it into a two-dimensional matrix.

[0035] The training module is used to extract features from the two-dimensional matrix using a convolutional neural network and organize them to obtain feature vectors; it sets up the dimensionality-reduced training model according to the characteristics of hourly meteorological data, and inputs the feature vectors into the set training model for training, so that the bidirectional long short-term memory network can achieve day-ahead load forecasting.

[0036] The pre-week load forecasting module is used to connect a bidirectional long short-term memory network to a gated loop unit based on a fully connected network. It inputs the abstract comprehensive data output by the bidirectional long short-term memory network into the gated loop unit to achieve pre-week load forecasting.

[0037] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to execute the multi-time-scale load forecasting method based on hybrid neural networks described above.

[0038] According to another aspect of the present invention, a processor is also provided, the processor being configured to run a program, wherein the program, when running, executes the multi-timescale load forecasting method based on hybrid neural networks as described above.

[0039] Compared with existing technologies, the present invention has the following advantages:

[0040] In this embodiment of the invention, considering the constraint relationship between the definite integral of the load curve and electricity consumption, as well as the influence of external factors on the load at different time scales, a hybrid neural network model is proposed. This model is used to correct the output of the pre-trained model. By extracting small-scale data into abstract comprehensive data through a multi-layer neural network and concatenating it with long-scale data to form a new feature vector, the relationship between data at different time scales and the load is fully explored, adding more prior knowledge to the model. Furthermore, the constraint relationship between the definite integral of the load curve and electricity consumption is incorporated into the loss function, enhancing the geometric correlation between various load elements and effectively reducing the impact of cumulative errors.

[0041] This invention takes into account the error of point prediction results and the physical meaning of the definite integral of the load curve, and can accurately predict the load at multiple time scales.

[0042] The load forecasting results of this invention at different time scales play an important supporting role in power system planning, dispatching, and the design of bidding systems in the electricity market. By extracting small-scale data into abstract comprehensive data through a multi-layer neural network and concatenating it with long-scale data to form a new feature vector, the relationship between data at different time scales and load is fully explored. This provides sufficient prior knowledge for load forecasting at multiple time scales and lays a solid foundation for the intelligent development of the power grid.

[0043] This invention fully considers the constraint relationship between the definite integral of the load curve and the electricity consumption, as well as the influence of external factors on the load at different time scales, effectively reducing the impact of cumulative errors when predicting long-term load. Attached Figure Description

[0044] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart of a multi-time-scale load forecasting method based on a hybrid neural network according to an embodiment of the present invention;

[0046] Figure 2 This is a structural diagram of the day-ahead load forecasting model network according to an embodiment of the present invention;

[0047] Figure 3 This is a network structure diagram of the pre-week load prediction model according to an embodiment of the present invention;

[0048] Figure 4 This is a schematic diagram of a daily-scale data fusion model according to an embodiment of the present invention;

[0049] Figure 5 This is a schematic diagram of the network output results of the day-ahead load forecasting model according to an embodiment of the present invention;

[0050] Figure 6 This is a schematic diagram of the network output results of the pre-week load prediction model according to an embodiment of the present invention. Detailed Implementation

[0051] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0052] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0053] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0054] Example 1

[0055] According to an embodiment of the present invention, an embodiment of a multi-timescale load forecasting method based on a hybrid neural network is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0056] like Figure 1 This is a flowchart of a multi-timescale load forecasting method based on a hybrid neural network according to an embodiment of the present invention, such as... Figure 1 and Figure 2 As shown, this multi-timescale load forecasting method includes the following steps:

[0057] Step S1: Obtain the original load data of the transformer, preprocess the original load data to obtain load data, use the load data as training data, and establish a pre-trained model based on a bidirectional long short-term memory network.

[0058] Step S2: Use a fully connected neural network to reduce the dimensionality of the output vector of the pre-trained model to obtain the dimensionality-reduced training model;

[0059] Step S3: Select historical meteorological data similar to the date to be predicted, and concatenate the hourly meteorological data in the historical meteorological data similar to the date to be predicted with the output of the fully connected neural network to convert it into a two-dimensional matrix;

[0060] Step S4: Use a convolutional neural network to extract features from the two-dimensional matrix and organize them to obtain feature vectors; set up the dimensionality-reduced training model according to the characteristics of hourly meteorological data, and input the feature vectors into the set training model for training, so that the bidirectional long short-term memory network can achieve day-ahead load forecasting.

[0061] Step S5: Connect the bidirectional long short-term memory network to the gated loop unit based on the fully connected network, and input the abstract comprehensive data output by the bidirectional long short-term memory network into the gated loop unit to realize the week-ahead load forecast.

[0062] As an optional embodiment, in step S1, the transformer load data obtained is the transformer load data at t-minute intervals within year T; where t < 60 and t = 15.

[0063] As an optional embodiment, in step S1, the preprocessing includes normalizing the original transformer load data, and the normalization expression is:

[0064]

[0065] In the formula, X is the normalized value. t X is the original load data for the transformer. min With X max These are the maximum and minimum loads from the original load data of the transformer. Normalized data is used as training data to build a pre-trained model based on a classic Long Short-Term Memory (LSTM) network.

[0066] As an optional embodiment, in step S1, the preprocessing further includes performing Lagrange interpolation on the missing load values, as shown below:

[0067] L' n (x)=y0l0(x)+y0l0(x)+···+y n l n (x) (2)

[0068] In the formula, l0(x), l1(x), ..., l n (x) is defined by x0, x1, ..., x n Let n be the interpolation basis function for the node.

[0069] As an optional implementation, a pre-trained model based on a bidirectional long short-term memory (BSSM) network can achieve load prediction on a 15-minute scale. Each basic unit of the BSSM network includes a forgetting gate. t Input gate i t and output gate o t， The relationships between the variables are shown below:

[0070]

[0071] In the formula, x t The input is the sample value for this layer, where t is the sample index value, and f is the input sample value. t Represents the forget gate unit vector, i t Represents the input gate unit vector, o t C represents the output gate unit vector. t Represents a cell state unit vector. h is the cell state transition state unit vector. t Let be the hidden layer unit vector, and w and b be the connection weights and biases between neurons in each unit, respectively. The input x in the forget gate... t With state memory unit C t-1 To achieve selective forgetting of the previous node, the input gate determines the current input x. t Saved to state memory unit C t-1 The quantity can be obtained through Selective output is achieved by controlling the influence of long-cycle load characteristics on the current output.

[0072] The optimizer Adam for Long Short-Term Memory Recurrent Neural Networks is:

[0073]

[0074] In the formula, Let represent the weights and biases of the network parameters at the nth iteration, respectively. and These are the accumulated second-order gradient momentum of the loss function during the first n iterations, β1 and β2 are the momentum gradient coefficients, and α is the learning rate. ε is the gradient correction amount, and ε is a small constant to prevent the denominator from being zero.

[0075] As an optional embodiment, step S2 uses a fully connected neural network to reduce the dimension of the output vector of the pre-trained model to the load data at 1-hour intervals, specifically including the following steps:

[0076] Based on the multi-input multi-output characteristics of fully connected neural networks, the time scale of the load data is changed, transforming the 15-minute interval load data into the 1-hour interval load data. The internal variables of the fully connected neural network layer are shown below:

[0077]

[0078] In the formula, n represents the number of input neurons, q represents the number of output neurons, p represents the number of hidden layer neurons, k is the sample index value, w and b are the connection weights and biases between neurons, respectively, and f(·) represents the activation function expression; hi h (k)with ho h (k) represent the output of the h-th hidden layer before and after the activation function for the k-th sample, respectively; yi o(k) with yo o (k) represents the output of the h-th output layer neuron before and after the activation function for the k-th sample.

[0079] The activation functions based on ReLU and LeakyReLU are shown below:

[0080]

[0081]

[0082] As an optional embodiment, the hybrid neural network in this invention needs to capture the nonlinear relationship between meteorological factors and load based on a two-dimensional matrix, and needs to couple the time-series characteristics of multiple variables. Therefore, step S3 specifically includes:

[0083] Step S31: Based on the principle of "nearer is larger and farther is smaller", select historical meteorological data similar to the day to be predicted from the training data as the meteorological data for the day to be predicted;

[0084] Step S32: Select hourly meteorological data from historical meteorological data similar to the predicted day, and select feature points of the hourly meteorological data. The feature points of the hourly meteorological data include: load L(i) at the hour, the highest temperature T(i) in one hour, and the highest air humidity H(i) in one hour.

[0085] Step S33: Based on the feature points of the hourly meteorological data, form a daily feature vector α(i) from the feature points, and then form a daily feature matrix A using a sliding window method from the daily feature vector α(i). Borrowing from the method of representing word vectors in natural language processing, transform the daily feature matrix A into a two-dimensional matrix:

[0086]

[0087] In the formula, α(i) = (L(i), T(i), H(i)). Specifically, the width of the sliding window is set to 8, the step size is set to 1, and the time axis is from left to right. The first column is time i, and the second column is time i+1.

[0088] As an optional embodiment, step S4 specifically includes:

[0089] Step S41: Set up the convolutional neural network.

[0090] Specifically, convolutional kernels are used to slide across a two-dimensional matrix to extract the relationships between variables. The output dimensions of the convolutional neural network layer are shown below:

[0091]

[0092] In the formula, `padding_size` represents the padding size, `w` represents the input sample width, `h` represents the input sample height, `w_out` represents the output sample width, `h_out` represents the output sample height, `f` represents the kernel size, and `s` represents the kernel stride. The convolutional neural network is configured with two layers, with the number of kernels set to 32 and 64 respectively. The kernel size is 3×3, and the pooling size in the pooling layer is set to 2. After two convolution and max pooling operations, a 2×2×64 three-dimensional vector array is obtained. This array is then flattened into a one-dimensional vector array of length 256 before being fed into a bidirectional long short-term memory network.

[0093] Step S42: Set up a bidirectional long short-term memory network.

[0094] Specifically, the units of the bidirectional long short-term memory network are the same as those of the long short-term memory network. The label value of the bidirectional long short-term memory network is set to the next day's load value, and the loss function is set to consider the error of the point prediction results and the influence of the definite integral of the load curve. The expression for the loss function is:

[0095]

[0096] In the formula, y represents the prediction result of the hybrid neural network at time t; t The actual load on the transformer at time t is obtained through measurement. The integral area of ​​the load curve between time t and time t-1; specifically, when making the week-ahead load forecast, t = 168.

[0097] Step S43: Use the convolutional neural network set in step S41 to extract features from the two-dimensional matrix and organize them to obtain feature vectors; input the feature vectors into the bidirectional long short-term memory network set in step S42 for training, so that the bidirectional long short-term memory network can achieve day-ahead load forecasting.

[0098] Specifically, when performing day-ahead load forecasting, T=24, the same Adam optimizer as the Long Short-Term Memory network is used to optimize the network. The day-ahead load forecasting result can be expressed as:

[0099] m(i)=F Bi-LSTM (n(i),m(i-1))(11)

[0100] In the formula, n(i) is the comprehensive abstract feature vector at time i after feature extraction by the convolutional neural network; m(i-1) is the load value at time i-1, where i = 1, 2, ... 24; F Bi-LSTM For bidirectional long short-term memory network input-output function, m(i) is the load at midnight of the next day and the load value at 24:00 of the same day.

[0101] As an optional embodiment, a gated recurrent unit (GRU) structure is added to the network structure based on the day-ahead load forecasting. The data processing method remains the same, but the difference lies in that the output of the bidirectional long short-term memory (LSTM) network is now an abstract, comprehensive data with temporal characteristics. The output of the bidirectional LSM network is then input into the GRU to achieve week-ahead load forecasting. Step S5 specifically includes the following steps:

[0102] Step S51: Taking into account the day type, following the principle of "nearer larger and farther smaller", select the weekly meteorological data that is close to the week to be predicted as the meteorological data of the week to be predicted. The meteorological feature set and feature description are shown in Table 1 below. The daily-scale meteorological data and load data are spliced ​​together based on a fully connected neural network.

[0103] Table 1. Meteorological Factor Feature Set and its Representation Method

[0104]

[0105] Step S52: Configure the gate circulation unit according to the characteristics of the meteorological data for the week to be predicted.

[0106] Specifically, the update gate of the gated recurrent unit selectively removes feature information from among numerous features and adds new feature information to the model in a timely manner according to the magnitude of the prediction error. The reset gate determines which load features to forget. The calculation formulas for the various model parameters of the gated recurrent unit are shown below:

[0107]

[0108] In the formula, z t Update the gate for the gate loop unit; W z Update the gate weights for the gate loop unit; r t Reset the door for the door circulation unit; W r Reset the weights of the gates in the gate loop unit; h t-1 The hidden layer state before time t; The product is the Hadamard product; σ() is the sigmoid activation function; tanh() is the hyperbolic tangent activation function, as shown below:

[0109]

[0110] Step S53: The daily meteorological data and load data are stitched together based on a fully connected neural network, such as... Figure 4 As shown, η t Let D be the load values ​​at all hourly times on day t of the following week, which is a one-dimensional vector of length 24; t-1 For day-scale data of day t; ε t This represents the abstract, comprehensive data processed by a bidirectional long short-term memory network on day t. Therefore, the overall process for predicting the load value for the following week can be expressed as:

[0111]

[0112] In the formula, n t (i) represents the comprehensive abstract feature vector of meteorological factors at the i-th hour of day t after feature extraction via a convolutional neural network; m t (i) represents the load value in the i-th hour of day t; ε t Let be a one-dimensional feature vector of length 48, representing an abstract feature vector of the load time series, which has no physical meaning. It uses the same loss function as the day-ahead load forecast, where T = 168 in the loss function. The load value on day t of the following week can be represented by the above variables as follows:

[0113] η t =F GRU (ε t D t ,η t-1 (15)

[0114] In the formula, F GRU D is a gate loop unit; t This is daily-scale data including weekday, weekday, holiday, daily high temperature, daily low temperature, daily high humidity, and daily low humidity on day t; η t-1 ε represents the load value at the previous time step. t This is an abstract, comprehensive set of data that contains the temporal characteristics of the load; finally, the load value for the following week is obtained through rolling prediction by the gate loop unit.

[0115] Example 2

[0116] One embodiment of the present invention will be described in detail to enable those skilled in the art to better understand the invention, specifically including the following steps:

[0117] 1) The distribution transformers in Portugal were selected as the main research object. Load data for 365 days in 2011 were selected as the raw data. The sampling time granularity of the data was 15 minutes. The relevant meteorological data, date type and load data were cleaned and normalized. Missing values ​​were supplemented based on Lagrange interpolation. The normalized data was used to train the Long Short-Term Memory Network as a pre-trained model.

[0118] 2) Use a fully connected neural network to reduce the dimensionality of the output vector of the pre-trained model to load data with an interval of 1 hour. Based on the load data output by the pre-trained model with an interval of 15 minutes, a fully connected network is used for comprehensive processing to reduce the dimensionality of the load vector to load data with an interval of 1 hour.

[0119] 3) Select historical meteorological data similar to the date to be predicted, and concatenate the hourly meteorological data with the output of the fully connected neural network to transform it into a two-dimensional matrix. Set the width of the sliding window to 8 and the step size to 1. The time axis is from left to right, with the first column representing time i and the second column representing time i+1.

[0120] 4) Feature extraction is performed on a two-dimensional matrix using a convolutional neural network. The model label is set to the next day's load value. The loss function considers the error of the point prediction result and the physical meaning of the definite integral of the load curve. The feature vector is input into a bidirectional long short-term memory network to train the model and achieve the day-ahead load prediction. The convolutional neural network is set to 2 layers, with the number of convolutional kernels set to 32 and 64 respectively, and the kernel size set to 3×3. The pool size in the pooling layer is set to 2. After two convolution and max pooling operations, a 2×2×64 three-dimensional vector array is obtained. Then, a flattening operation is performed to change it into a one-dimensional vector array of length 256. The one-dimensional vector array of length 256 is input into the bidirectional long short-term memory network. The bidirectional long short-term memory network has 2 layers, a learning rate of 0.001, and Dropout is set to 0.3. The prediction result is as follows. Figure 5 As shown in Table 2 below, the Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE) are used as evaluation criteria to compare the patented model with various prediction models.

[0121] Table 2 Calculation results of evaluation indicators for each prediction model

[0122]

[0123] To verify the impact of considering the error of the point prediction results and the physical meaning of the definite integral of the load curve on the accuracy of day-ahead load forecasting, based on the same hybrid neural network structure, models with and without definite integral constraints were compared. The results are shown in the table below:

[0124] Table 3 Comparison of experimental results

[0125]

[0126] 5) Daily timescale meteorological factors are connected to gated recurrent units (GRUs) through fully connected layers. The abstract, comprehensive data output from the bidirectional long short-term memory (LSTM) network is directly input into the GRU. The model label is set to the next week's load value. The loss function considers the error of the point prediction results and the physical meaning of the definite integral of the load curve. The model can ultimately achieve weekly load prediction. The convolutional neural network is set to 2 layers, with the number of convolutional kernels set to 32 and 64 respectively. The kernel size is 3×3, and the pool size in the pooling layer is set to 2. After two convolutions and max pooling operations, a 2×2×64 three-dimensional vector array is obtained. This array is then flattened into a one-dimensional vector array of length 256. This 256-length one-dimensional vector array is input into the bidirectional LSM network, which has 2 layers and 1 GRU. The learning rate is 0.001, and Dropout is set to 0.3. The prediction results are as follows. Figure 6 As shown in Table 4 below, the Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE) are used as evaluation criteria to compare the patented model with various prediction models.

[0127] Table 4. Calculation results of evaluation indicators for each prediction model

[0128]

[0129] To verify the impact of the error in the point prediction results and the physical meaning of the definite integral of the load curve on the prediction accuracy a week in advance, a comparison was made between the model with and without definite integral constraints based on the same hybrid neural network structure. The results are shown in Table 5 below. As can be seen from Table 5, the load prediction is more accurate after adding definite integral constraints.

[0130] Table 5 Comparison of experimental results

[0131]

[0132] Example 3

[0133] According to another aspect of the present invention, a multi-time-scale load forecasting system based on a hybrid neural network is also provided. The multi-time-scale load forecasting system applies the above-described multi-time-scale load forecasting method based on a hybrid neural network. The multi-time-scale load forecasting system includes:

[0134] The pre-trained model module is used to acquire the original load data of the transformer, preprocess the original load data to obtain load data, and establish a pre-trained model based on a bidirectional long short-term memory network.

[0135] The dimension reduction module is used to reduce the dimension of the output vector of the pre-trained model using a fully connected neural network to obtain the dimension-reduced training model.

[0136] The two-dimensional matrix module is used to select historical meteorological data similar to the date to be predicted, and to concatenate the hourly meteorological data in the historical meteorological data similar to the date to be predicted with the output of the fully connected neural network to transform it into a two-dimensional matrix.

[0137] The training module is used to extract features from the two-dimensional matrix using a convolutional neural network and organize them to obtain feature vectors; it sets up the dimensionality-reduced training model according to the characteristics of hourly meteorological data, and inputs the feature vectors into the set training model for training, so that the bidirectional long short-term memory network can achieve day-ahead load forecasting.

[0138] The pre-week load forecasting module is used to connect a bidirectional long short-term memory network to a gated loop unit based on a fully connected network. It inputs the abstract comprehensive data output by the bidirectional long short-term memory network into the gated loop unit to achieve pre-week load forecasting.

[0139] This invention is not limited to the specific embodiments described above. The above are merely preferred embodiments of this invention and are not intended to limit the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

[0140] Example 4: According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored program, wherein, when the program is running, it controls the device where the computer-readable storage medium is located to execute any of the above-described multi-timescale load forecasting methods based on hybrid neural networks.

[0141] Optionally, in this embodiment, the computer-readable storage medium may be located in any computer terminal in a group of computer terminals in a computer network, or in any mobile terminal in a group of mobile terminals, and the computer-readable storage medium includes a stored program.

[0142] Optionally, during program execution, the device containing the computer-readable storage medium performs the following functions: acquiring the original load data of the transformer; preprocessing the original load data to obtain load data; and establishing a pre-trained model based on a bidirectional long short-term memory network; using a fully connected neural network to reduce the dimensionality of the output vector of the pre-trained model to obtain a dimensionality-reduced training model; selecting historical meteorological data similar to the date to be predicted; concatenating the hourly meteorological data from the historical meteorological data similar to the prediction date with the output of the fully connected neural network to convert it into a two-dimensional matrix; using a convolutional neural network to extract features from the two-dimensional matrix and organizing them to obtain feature vectors; setting the dimensionality-reduced training model according to the characteristics of the hourly meteorological data, and inputting the feature vectors into the set training model for training, so that the bidirectional long short-term memory network can achieve day-ahead load prediction; connecting the bidirectional long short-term memory network with a gated loop unit based on a fully connected network, and inputting the abstract comprehensive data output of the bidirectional long short-term memory network into the gated loop unit to achieve week-ahead load prediction.

[0143] Example 5

[0144] According to another aspect of the present invention, a processor is also provided for running a program, wherein the program executes any of the above-described multi-timescale load forecasting methods based on hybrid neural networks.

[0145] This invention provides an apparatus including a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of a multi-timescale load forecasting method based on a hybrid neural network.

[0146] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0147] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0148] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The system embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interface, and the indirect coupling or communication connection of units or modules may be electrical or other forms.

[0149] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0150] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0151] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0152] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multi-time-scale load forecasting method based on hybrid neural networks, characterized in that, include: The original load data of the transformer is obtained, the original load data is preprocessed to obtain load data, and a pre-trained model is established based on a bidirectional long short-term memory network. The output vector of the pre-trained model is reduced in dimensionality using a fully connected neural network to obtain the dimensionality-reduced training model. Select historical meteorological data similar to the date to be predicted, and then concatenate the hourly meteorological data from the historical meteorological data similar to the date to be predicted with the output of the fully connected neural network to transform it into a two-dimensional matrix. A convolutional neural network is used to extract features from the two-dimensional matrix and organize them to obtain feature vectors. The dimensionality-reduced training model is set according to the characteristics of hourly meteorological data, and the feature vectors are input into the set training model for training, so that the bidirectional long short-term memory network can achieve day-ahead load forecasting. By connecting a bidirectional long short-term memory network with a gated loop unit based on a fully connected network, the abstract comprehensive data output by the bidirectional long short-term memory network is input into the gated loop unit to achieve week-ahead load forecasting. The method of connecting a bidirectional long short-term memory network (LSTM) with a gated recurrent unit (GRU) based on a fully connected network, enabling the abstract and comprehensive data output by the LSM to achieve week-ahead load forecasting, includes: Weekly meteorological data close to the week to be predicted is selected as the meteorological data for the week to be predicted. Daily timescale meteorological factors are connected to gated recurrent units through fully connected layers, that is, the abstract comprehensive data output by the bidirectional long short-term memory network is directly input into the gated recurrent unit. The gate circulation unit is configured based on the characteristics of the meteorological data for the week to be predicted; The gate loop unit is configured to output the load value on day t of the next week. : In the above formula, It is a gate loop unit. This is daily-scale data for day t, which includes weekday, weekday, holiday, daily high temperature, daily low temperature, daily high humidity, and daily low humidity. The load value at the previous time point. This represents the abstract, comprehensive data processed by the bidirectional long short-term memory network on day t. The load value for the following week is obtained through rolling prediction using the gate loop unit.

2. The multi-time-scale load forecasting method based on hybrid neural networks according to claim 1, characterized in that, The obtained transformer load data is the transformer load data at t-minute intervals within year T.

3. The multi-time-scale load forecasting method based on hybrid neural networks according to claim 1, characterized in that, Using a fully connected neural network reduces the dimension of the output vector of the pre-trained model to the load data at 1-hour intervals.

4. The multi-time-scale load forecasting method based on hybrid neural networks according to claim 1, characterized in that, The hourly meteorological data is concatenated with the output of a fully connected neural network and transformed into a two-dimensional matrix, including: Based on the characteristic points of hourly meteorological data, the characteristic points are used to form a daily feature vector. Among them, the characteristic points of hourly meteorological data include: hourly load. The highest temperature in one hour The highest value of air humidity within one hour ; Daily feature vectors The daily feature matrix A is constructed using a sliding window method; Transform the daily feature matrix A into a two-dimensional matrix: In the above formula, .

5. The multi-time-scale load forecasting method based on hybrid neural networks according to claim 1, characterized in that, Based on the characteristics of hourly meteorological data, the training model after dimensionality reduction is configured as follows: the label value of the bidirectional long short-term memory network in the pre-trained model is set to the next day's load value, and the loss function is set to consider the error of the point prediction results and the influence of the definite integral of the load curve.

6. The multi-time-scale load forecasting method based on hybrid neural networks according to claim 1, characterized in that, The expression is: In the formula, This is a comprehensive abstract feature vector of meteorological factors at the i-th hour of day t after feature extraction by a convolutional neural network. Let i be the load value for the i-th hour on day t. This is the input / output function for a bidirectional long short-term memory network.

7. A multi-time-scale load forecasting system based on a hybrid neural network, characterized in that, include: The pre-trained model module is used to acquire the original load data of the transformer, preprocess the original load data to obtain load data, and establish a pre-trained model based on a bidirectional long short-term memory network. The dimension reduction module is used to reduce the dimension of the output vector of the pre-trained model using a fully connected neural network to obtain the dimension-reduced training model. The two-dimensional matrix module is used to select historical meteorological data similar to the date to be predicted, and to concatenate the hourly meteorological data in the historical meteorological data similar to the date to be predicted with the output of the fully connected neural network to transform it into a two-dimensional matrix. The training module is used to extract features from the two-dimensional matrix using a convolutional neural network and organize them to obtain feature vectors; it sets up the dimensionality-reduced training model according to the characteristics of hourly meteorological data, and inputs the feature vectors into the set training model for training, so that the bidirectional long short-term memory network can achieve day-ahead load forecasting. The pre-week load forecasting module is used to connect a bidirectional long short-term memory network to a gated loop unit based on a fully connected network, and input the abstract comprehensive data output by the bidirectional long short-term memory network into the gated loop unit to realize pre-week load forecasting. The method of connecting a bidirectional long short-term memory network (LSTM) with a gated recurrent unit (GRU) based on a fully connected network, enabling the abstract and comprehensive data output by the LSM to achieve week-ahead load forecasting, includes: Weekly meteorological data close to the week to be predicted is selected as the meteorological data for the week to be predicted. Daily timescale meteorological factors are connected to gated recurrent units through fully connected layers, that is, the abstract comprehensive data output by the bidirectional long short-term memory network is directly input into the gated recurrent unit. The gate circulation unit is configured based on the characteristics of the meteorological data for the week to be predicted; The gate loop unit is configured to output the load value on day t of the next week. : In the above formula, It is a gate loop unit. This is daily-scale data for day t, which includes weekday, weekday, holiday, daily high temperature, daily low temperature, daily high humidity, and daily low humidity. The load value at the previous time point. This represents the abstract, comprehensive data processed by the bidirectional long short-term memory network on day t. The load value for the following week is obtained through rolling prediction using the gate loop unit.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device containing the computer-readable storage medium to perform the multi-timescale load forecasting method based on any one of claims 1 to 6.

9. A processor, characterized in that, The processor is used to run a program, wherein the program executes the multi-timescale load forecasting method based on a hybrid neural network as described in any one of claims 1 to 6.

Images