Power load prediction method based on decomposition network

Abstract

The present application relates to the field of power load prediction, and more particularly to a power load prediction method based on a decomposition network, comprising: collecting historical load data and weather data of a user side as original data; preprocessing the original data; constructing a prediction model and training; inputting historical load data and weather data of a day to be tested into the trained prediction model to obtain a power load prediction value of the day to be tested; in the prediction model, the prediction model comprises a decomposition network, a convolution layer and a multilayer perceptron; a decomposition module in the decomposition network comprises a front-back cutting unit and a local correlation information extraction unit; the preprocessed original data is taken as an input sequence, the front-back cutting unit cuts the input sequence into a front sequence and a rear sequence, and then extracts vector data in the sequence through two local correlation information extraction units to output a subsequence; the decomposition network finally obtains a plurality of groups of subsequences; and finally, a load prediction value is output through the convolution and the multilayer perceptron. The present application has an accurate prediction effect.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting, and more particularly to a power load forecasting method based on decomposition networks. Background Technology

[0002] With the continuous development of society, the demand for energy from individuals and businesses is increasing. Among the many types of energy, electricity, as a clean energy source, accounts for a large proportion of social energy consumption. A safe, stable, and economical power system is the foundation for maintaining social order and stability, and power load forecasting is an important component of the power system. Short-term power load forecasting refers to the forecasting of load within the next few hours, day, or week, and it has a positive impact on reducing energy consumption and optimizing resource allocation.

[0003] Current load forecasting methods mainly include time series forecasting, regression analysis forecasting, and neural network forecasting. Time series forecasting identifies load change trends in time series data, enabling effective prediction of future loads. However, it doesn't consider the influence of external factors, and significant changes in these factors often lead to large deviations in the prediction results. Regression analysis forecasting uses one or more independent variables to predict load trends, but its short-term load predictions are often unsatisfactory due to the lack of learning capabilities in its models. Neural network forecasting uses historical data as input, processes it through hidden layers, and outputs predicted values. The error between the predicted and actual values ​​is used as the objective function. The network parameters are updated by continuously reducing the error until a suitable error is achieved, at which point the parameter values ​​are fixed to establish the final prediction model. However, neural network forecasting requires a large amount of data and has relatively high computational costs.

[0004] Therefore, in order to overcome the above-mentioned shortcomings, providing a power load forecasting method based on decomposition networks has become an urgent problem to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a power load forecasting method based on decomposition networks, which has accurate forecasting results.

[0006] To solve the above technical problems, the technical solution of the present invention is: a power load forecasting method based on decomposition networks, comprising:

[0007] Step 1: Collect historical load data and meteorological data from the user side as raw data;

[0008] Step 2: Preprocess the raw data;

[0009] Step 3: Build and train the prediction model:

[0010] The prediction model includes a decomposition network, convolutional layers, and a multilayer perceptron; the decomposition network has a multilayer structure, and each layer of the decomposition network includes several decomposition modules; the decomposition modules include front and back segmentation units and local correlation information extraction units.

[0011] The preprocessed raw data is used as the input sequence. The pre- and post-segmentation unit divides the input sequence into a pre-sequence and a post-sequence. The pre- and post-sequences are then processed by two local correlation information extraction units to extract vector data from the sequence, resulting in two sub-sequences. The output sub-sequences are then fed into the decomposition modules of the next layer until the final layer of the decomposition network is reached, resulting in several sets of sub-sequences. Each set of sub-sequences is then passed through a convolutional layer to obtain several values. Finally, these values ​​are processed by a multilayer perceptron to output the load prediction value.

[0012] Step 4: Input the historical load data and meteorological data of the day to be measured into the trained prediction model to obtain the predicted power load value of the day to be measured.

[0013] Furthermore, in step 1, the load data is the historical average total load power per day and hour within the target area; the meteorological data includes the average temperature and average humidity per day and hour.

[0014] Furthermore, step 2 specifically includes:

[0015] Identify and correct outliers in the original data;

[0016] Impute missing values ​​in the original data.

[0017] Each type of original data that has been corrected and filled is normalized to eliminate the influence of feature dimensions and improve the accuracy of the model.

[0018] Furthermore, the method for judging and correcting outliers in the original data is as follows: the 3sigma criterion is used to determine whether it is an outlier point in a day. For outlier points, values ​​are assigned and corrected based on the data at two times before and after the same day and the data at the same time on the two days before and after.

[0019] Furthermore, for missing data points in the original data: if the number of missing points is greater than or equal to half the number of records in a day, then the data for the current day is deleted; if the number of missing points is less than half the number of records in a day, then the average of the data before and after the current missing point is used to fill in the missing points.

[0020] Furthermore, the normalization method adopted is the min-max normalization method.

[0021] Furthermore, the local correlation information extraction unit adopts a masking mechanism, which makes any vector in the input sequence of the unit only correlated with itself and vectors within a certain range before it. The masking mechanism is conducive to the extraction of effective information that conforms to the temporal logic, while reducing the computational load of the network.

[0022] Furthermore, in step 3, each day is divided into T time points with equal time intervals, and the feature dimension in the original data is P; the feature of each time point is represented as a vector with dimension P, that is, there are T vectors per day; data from N consecutive days is taken for training, that is, there are T*N vectors, which are used as the input sequence, that is, the length of the input sequence is T*N and the dimension is P.

[0023] The input sequence of length T*N and dimension P is decomposed through several layers of decomposition modules to extract local correlation information, resulting in several subsequences of the same length. Each subsequence is then passed through a convolutional layer to obtain several values. Finally, these values ​​are passed through a multilayer perceptron to output T values, which represent the load forecast values ​​at each time point of the next day.

[0024] In this module, the first-level decomposition unit splits the input sequence of length T*N and dimension P into a front sequence and a back sequence, with lengths of T*N and P respectively. The dimension is P;

[0025] The decomposition module has two local association information extraction units used to extract vector data from the preceding and following sequences. These local association information extraction units employ a masking mechanism, specifically:

[0026] The local association information extraction unit considers the preceding or following sequence as having a row height of Given a matrix I with column width P, multiply I by three parameter matrices W of size P*P. q W k W v We obtain the query matrix Q, the key matrix K, and the value matrix V, which can be represented using matrix multiplication as follows:

[0027] Q = IW q

[0028] K = IW k

[0029] V = IW v

[0030] Multiply the query matrix Q by the transpose of the key matrix K to obtain the association matrix A;

[0031] Define a mask matrix E, whose row height or column width is the same as the correlation matrix A. The mask matrix consists of (-10^6, 0). The number of 0 values ​​in the mask matrix E increases row by row and eventually becomes fixed at a certain value, so that any vector in the input sequence is correlated only with vectors within a certain adjacent range. The -10^6 value in the mask matrix E ensures that any vector in the input sequence is correlated only with itself and the vectors preceding it.

[0032] When the mask matrix E is added to the correlation matrix A and a softmax operation is performed, the normalized correlation matrix A′ is obtained; the value in the correlation matrix A′ corresponding to -10^6 in the mask matrix E is 0, and the values ​​in the other positions are normalized values.

[0033] Finally, the correlation matrix A′ is multiplied by the value matrix V to obtain the final output matrix O, with a row height of . The column width is P, and each element in each row represents a vector at the corresponding position in the output sequence. That is, the output of any local correlation information extraction unit in the first-level decomposition module is a vector of length P. A subsequence of dimension P;

[0034] After processing by the B-layer decomposition module, the final result is 2. B Grouped subsequences, each subsequence having a length of [length missing]. The dimension is P;

[0035] Each subsequence is passed through a convolutional layer to output 2. B One value;

[0036] Finally, these 2 B The data is processed by a multilayer perceptron, which outputs T values, representing the load forecast value at each time point on the day to be measured.

[0037] Furthermore, step 3 also includes a step of testing the prediction model; specifically:

[0038] The loss function, Loss, represents the magnitude of the error between the predicted and actual values. A smaller loss function value indicates a smaller error between the predicted and actual values. The mean squared error is used to represent the loss function, and the calculation formula is as follows:

[0039]

[0040] Among them, y i This represents the actual load value at time i. Let $\frac{i}{i}$ represent the predicted load value at time $i$, and $\frac{i}{i}$ represent the value of the loss function. Through the gradient descent algorithm, the parameters in the model will be continuously updated iteratively, causing the $\frac{i}{i}$ value to continuously decrease. When the $\frac{i}{i}$ value hardly changes anymore, it means that the prediction model has converged, the training process is over, and the parameters of the model are fixed.

[0041] Furthermore, when collecting raw data, the average total load power per hour in the target area is used as historical load data, and the average temperature and humidity per hour are used as meteorological data; the data predicted by the prediction model is the hourly load prediction value for the day to be measured.

[0042] The present invention has the following beneficial effects:

[0043] I. This invention uses a decomposition network for extracting local correlation information. This network consists of multiple layers of decomposition modules for extracting local correlation information. As the number of decomposition modules in the decomposition network increases, deeper temporal information will be extracted. Through this structure, the decomposition network completes the transition from overall information extraction to local information extraction, so that the output prediction value takes into account both the long-term and short-term relationships with the input data, thereby achieving accurate prediction of the prediction model.

[0044] Second, the decomposition module includes a front and rear segmentation unit and a local correlation information extraction unit. The special masking mechanism used in the local correlation information extraction unit makes any vector in the input sequence of this unit only calculate its correlation with itself and the vectors within a certain range before it. This not only reduces the computational load of the network, but also helps to extract effective information that conforms to the temporal logic more quickly, thereby accelerating the convergence of the network. Attached Figure Description

[0045] Figure 1 This is a flowchart of the prediction method of the present invention;

[0046] Figure 2 This is a schematic diagram of the decomposed network in an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of the decomposition module in an embodiment of the present invention;

[0048] Figure 4 This is a schematic diagram of the mask matrix in an embodiment of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0050] Please refer to Figure 1This embodiment provides a flowchart for load forecasting using a decomposition network to extract local correlation information. First, historical load data and meteorological data from the user side are collected and preprocessed. Then, the preprocessed data is used to train a decomposition network to extract local correlation information, constructing a prediction model. Finally, the load situation of the target area for each hour the following day is predicted based on the model.

[0051] This invention relates to a power load forecasting method based on decomposition networks, comprising the following steps:

[0052] Step 1: Collect historical load data and meteorological data from the user side as raw data; specifically:

[0053] Collect the average total load power per day and hour within the target area over a period of time, and use it as the raw historical load data; collect the average temperature and average humidity per day and hour within the target area as the raw meteorological data.

[0054] Step 2: Preprocess the raw data;

[0055] Raw historical load and meteorological data may contain anomalies or missing data, which often leads to a decrease in the accuracy of model predictions. Therefore, preprocessing of outliers and missing values ​​is necessary; specifically, this includes: identifying and correcting outliers in the raw data; imputing missing values ​​in the raw data; and normalizing each type of raw data after correction and imputation.

[0056] For outliers in the raw data, we first use the 3 sigma criterion to determine whether they are outliers of the day. The formula is as follows:

[0057]

[0058] Among them, v n,t This represents the value at time t on day n. Let σ represent the mean of all times on day n, and let σ represent the standard deviation of all times on day n. For points that satisfy the above formula, they are identified as outliers. Outlier values ​​are then corrected based on data from the two times before and after the same day, as well as data from the two times before and after the same day. The formula for correcting outliers is as follows:

[0059]

[0060] Among them, v n,t ′ represents the corrected value at time t on day n, v n+1,t and v n-1,t The values ​​of v at time t on day (n+1) and day (n-1) are respectively. n,t+1 and v n,t-1These are the values ​​at time t+1 and t-1 on day n, respectively. α1 and α2 are weighting coefficients, and their sum should equal 1.

[0061] For missing data points in the original data, if the number of missing points is greater than or equal to half the number of records in a day, then the data for that day is deleted. If the number of missing points is less than half the number of records in a day, then the average of the values ​​at the two times before and after the missing point is used to fill in the missing points. The formula for filling in missing values ​​is as follows:

[0062]

[0063] Among them, v n,t ' is the value at time t on day n after filling, v n,t+1 and v n,t-1 The values ​​of v at time t+1 and t-1 on day n are respectively. n,t+2 and v n,t-2 These are the values ​​at time t+2 and t-2 on day n, respectively. β1 and β2 are weighting coefficients, and their sum should equal 1.

[0064] After correcting and imputing outliers and missing values, it is necessary to normalize the three sets of data—temperature, humidity, and load power—each day to eliminate the influence of different dimensions and improve the model's accuracy. The normalization method used is min-max normalization, and the normalization formula for any set of data is as follows:

[0065]

[0066] in, R represents the normalized value at time t on day n. n,t MaxR represents the original value at time t on day n. n and minR n Let represent the maximum and minimum values ​​at all times on day n, respectively.

[0067] Step 3: Build and train the prediction model:

[0068] like Figure 2 and 3 As shown, a prediction model based on a decomposition network for extracting local correlation information is constructed. The prediction model includes a decomposition network, convolutional layers, and a multilayer perceptron; the decomposition network has a multilayer structure, and each layer of the decomposition network includes several decomposition modules; the decomposition modules include front and rear segmentation units and local correlation information extraction units;

[0069] The preprocessed raw data is used as the input sequence. The pre- and post-segmentation unit divides the input sequence into a pre-sequence and a post-sequence. The pre- and post-sequences are then processed by two local correlation information extraction units to extract vector data from the sequence, resulting in two sub-sequences. The output sub-sequences are then fed into the decomposition modules of the next layer until the final layer of the decomposition network is reached, resulting in several sets of sub-sequences. Each set of sub-sequences is then passed through a convolutional layer to obtain several values. Finally, these values ​​are processed by a multilayer perceptron to output the load prediction value.

[0070] The day is divided into 24 time points, T = 24. The temperature, humidity, and load power of each hour can be represented as a vector of dimension P = 3, meaning there are 24 vectors per day. Data from N = 4 consecutive days is used for training, resulting in 96 vectors. These are treated as sequences, meaning the input sequence has a length of 96 and a dimension of 3.

[0071] First, the input sequence is passed through the decomposition module (LAD-block) of the first layer to extract local correlation information, resulting in two subsequences with a length of 48 and a dimension of 3.

[0072] The structure diagram of the decomposition module (LAD-block) for extracting local correlation information is as follows: Figure 3 As shown, this module consists of a front-to-back segmentation unit and a local correlation information extraction unit. Taking the first layer LAD-block as an example, its front-to-back segmentation unit divides the input sequence (length 96, dimension 3) of this module into a front sequence and a back sequence (length 48, dimension 3); then, the front sequence and the back sequence are respectively fed into two local correlation information extraction units. The local correlation information extraction unit treats the front sequence or the back sequence as a matrix I with a row height of 48 and a column width of 3, and multiplies I by three parameter matrices W of size 3*3. q W k W v We obtain the query matrix Q, the key matrix K, and the value matrix V, which can be represented using matrix multiplication as follows:

[0073] Q = IW q

[0074] K = IW k

[0075] V = IW v

[0076] The correlation matrix A can be obtained by multiplying the query matrix Q by the transpose of the key matrix K.

[0077] Based on the characteristics of time series prediction tasks, a mask matrix E is defined and added to the correlation matrix A; where the mask matrix diagram is shown below. Figure 4As shown; the mask matrix consists of (-10^6, 0), and only sufficiently large negative values ​​become 0 after passing through softmax; Figure 4 The values ​​of the padding portion are set to -10^6, and the values ​​of the white portion are set to 0. The row height or column width of the mask matrix E is the same as that of the correlation matrix A. The inverted triangle design in the diagram ensures that any vector in the input sequence is correlated only with itself and the vectors preceding it. Figure 4 The design, which gradually increases the number of white portions and eventually fixes them at a certain value, ensures that any vector in the input sequence is only correlated with vectors within a certain adjacent range. The local correlation information extraction unit employs a special masking mechanism, ensuring that any vector in the input sequence is only correlated with itself and vectors within a certain preceding range. This operation facilitates the extraction of effective information conforming to temporal logic while reducing the computational load of the network.

[0078] When the mask matrix E is added to the correlation matrix A and a softmax operation is performed, a normalized correlation matrix A′ is obtained. The values ​​in A′ corresponding to the positions filled in the mask matrix E are 0, and the values ​​at the remaining positions are normalized values.

[0079] Finally, A′ is multiplied by the value matrix V to obtain the final output matrix O. The output matrix has a row height of 48 and a column width of 3. Each element in each row represents the vector at the corresponding position in the output sequence. That is, the output of any local association information extraction unit in the first layer LAD-block is a subsequence with a length of 48 and a dimension of 3.

[0080] After processing by a B=5 layer LAD-block, 32 subsequences are obtained, each with a length of 3 and a dimension of 3. Each subsequence is then passed through a convolutional layer with a 3*3 kernel, outputting 32 values. Finally, these 32 values ​​are passed through a multilayer perceptron, outputting 24 values ​​representing the hourly load forecast for the target day (i.e., the following day).

[0081] Step 3 also includes testing the prediction model; specifically:

[0082] The loss function, Loss, represents the magnitude of the error between the predicted and actual values. A smaller loss function value indicates a smaller error between the predicted and actual values. The mean squared error is used to represent the loss function, and the calculation formula is as follows:

[0083]

[0084] Among them, y i This represents the actual load value at time i. Let $\frac{i}{i}$ represent the predicted load value at time $i$, and $\frac{i}{i}$ represent the value of the loss function. Through the gradient descent algorithm, the parameters in the model will be continuously updated iteratively, causing the $\frac{i}{i}$ value to continuously decrease. When the $\frac{i}{i}$ value hardly changes anymore, it means that the prediction model has converged, the training process is over, and the parameters of the model are fixed.

[0085] Step 4: Predict the load situation for the next day using the model:

[0086] By using the historical load data, temperature data, and humidity data from the previous four days as input to the already trained model (i.e., the input is a sequence of length 96 and dimension 3), the hourly load forecast for the next day can be output.

[0087] This invention combines users' historical load data and other relevant data to construct and train a model that can accurately predict short-term power load.

[0088] All parts not covered in this invention are the same as or implemented using existing technologies.

[0089] The above description, in conjunction with specific embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for power load forecasting based on a decomposition network, characterized in that: The method steps include: Step 1: Collecting user-side historical load data and meteorological data as original data; Step 2: Preprocessing the original data; Step 3: Building a prediction model and training: The prediction model includes a decomposition network, a convolution layer and a multilayer perceptron; the decomposition network is a multi-layer structure, and each layer of the decomposition network includes a plurality of decomposition modules; the decomposition module includes a front-back split unit and a local correlation information extraction unit; The preprocessed original data is used as an input sequence, the front-back split unit splits the input sequence into a front sequence and a rear sequence, and the front sequence and the rear sequence are extracted through two local correlation information extraction units to obtain two sub-sequences; the output sub-sequences are input into each decomposition module of the next layer respectively until the final layer of the decomposition network is reached to obtain a plurality of groups of sub-sequences; each group of final sub-sequences is passed through a convolution layer to obtain a plurality of numerical values; finally, the plurality of numerical values are output through a multilayer perceptron to obtain a load prediction value; Step 4: Inputting the historical load data and meteorological data of the day before the to-be-tested day into the trained prediction model to obtain the power load prediction value of the to-be-tested day; The local correlation information extraction unit adopts a mask mechanism, so that any vector in the input sequence of the unit only calculates the correlation with itself and the vectors within a certain range before it. 2.The decomposition network-based power load forecasting method of claim 1, wherein: In step 1, the load data is the average total load power of each hour of each day in the target area; the meteorological data includes the average temperature and the average humidity of each hour of each day. 3.The decomposition network-based power load forecasting method of claim 1, wherein: Step 2 specifically includes: Judging the abnormal values in the original data and correcting the abnormal values; Filling in the missing values in the original data; Normalizing each type of original data after correction and filling. 4.The method of claim 3, wherein: The method for judging the abnormal values in the original data and correcting the abnormal values is: judging whether it is an abnormal value point in a day by the 3sigma criterion, and assigning a value to the abnormal value point according to the data of two time points before and after it and the data of the same time point of two days before and after it. 5.The decomposition network-based power load forecasting method of claim 3, wherein: For the missing value points in the original data: if the number of missing points is greater than or equal to half of the number of recording points in a day, the current day data is deleted; if the number of missing points is less than half of the number of recording points in a day, the average of the data of the time points before and after the missing point is taken to fill in the missing point. 6.The decomposition network-based power load forecasting method of claim 3, wherein: The normalization method adopts the min-max normalization method. 7.The decomposition network-based power load forecasting method of claim 1, wherein: In step 3, each day is divided into T time points with equal time intervals, and the feature dimension of the original data is P; the feature of each time point is represented as a vector with a dimension of P, i.e. there are T groups of vectors per day; taking the data of continuous N days for training, i.e. there are T*N groups of vectors, which are used as input sequences, i.e. the length of the input sequence is T*N and the dimension is P; The input sequence with a length of T*N and a dimension of P is passed through a plurality of decomposition modules for extracting local correlation information to obtain a plurality of groups of sub-sequences with the same length; each group of sub-sequences is passed through a convolution layer to obtain a plurality of numerical values; finally, the plurality of numerical values are output through a multilayer perceptron to output T values, which represent the load prediction values of each time point of the next day respectively; Wherein, the front and back cutting units of the first layer decomposition module cut the input sequence with length T*N and dimension P into a front sequence and a back sequence by cutting in half from front to back, and the length of the front sequence and the back sequence is dimension P; The two local correlation information extraction units of the decomposition module are used to extract vector data in the preceding sequence and the following sequence; the local correlation information extraction unit adopts a mask mechanism, and specifically is: The local correlation information extraction unit regards the pre-sequence or post-sequence as a matrix I with a row height of and a column width of P, and multiplies I by three parameter matrices W q , W k , and W v of size P*P respectively to obtain a query matrix Q, a key matrix K, and a value matrix V, which are expressed by matrix multiplication as follows: Q = IW q K = IW k V = IW v The query matrix Q is multiplied by the transpose of the key matrix K to obtain the correlation matrix A; A mask matrix E is defined, the row height or column width of the mask matrix E is the same as that of the correlation matrix A, the mask matrix is composed of (-10^6, 0), the number of 0 values in the mask matrix E increases row by row and is finally fixed to a certain value, so that any vector in the input sequence only calculates the correlation with the vectors within a certain adjacent range; the -10^6 value in the mask matrix E makes any vector in the input sequence only calculate the correlation with itself and the previous vector; When the mask matrix E is added to the correlation matrix A and a softmax operation is performed once, the normalized correlation matrix A' is obtained; the value corresponding to the -10^6 in the mask matrix E in the correlation matrix A' is 0, and the values of the remaining positions are normal normalized values; Finally, the relevance matrix A' is multiplied by the value matrix V to obtain the final output matrix O, the row height of the output matrix O is and the column width is P, and the elements of each row represent the vector at the corresponding position in the output sequence, that is, the output of any local relevance information extraction unit in the first layer decomposition module is a subsequence with a length of P dimensions; After the processing of the B-layer decomposition module, the final 2 B group subsequences, each group subsequence having a length of dimension P; Each group of sub-sequences is respectively passed through a convolution layer with a kernel size of 3*3, and two values are obtained B from the output. Finally, the two values are outputted by a multi-layer perception, and T values are outputted, which represent the load prediction values of each time point of the day to be measured. B Finally, the two values are outputted by a multi-layer perception, and T values are outputted, which represent the load prediction values of each time point of the day to be measured. 8.The decomposition network-based power load forecasting method of claim 1, wherein: The step 3 further includes a step of testing the prediction model; specifically: The loss function Loss is used to represent the error size between the predicted value and the true value, the smaller the value of the loss function, the smaller the error between the predicted value and the true value; the mean square error is used to represent the loss function, and the calculation formula is as follows: wherein y i represents the real load value at the i th moment, represents the predicted load value at the i th moment, Loss represents the value of the loss function; through the gradient descent algorithm, the parameters in the model will be iteratively updated, and the Loss value will be continuously reduced. When the Loss value hardly changes anymore, it means that the prediction model has converged, the training process is completed, and the parameters of the model are fixed. 9.The decomposition network-based power load forecasting method of claim 1, wherein: When collecting the original data, the average total load power in the target area per day per hour is taken as the historical load data, and the average temperature and average humidity per day per hour are taken as the meteorological data; the data predicted by the prediction model is the load prediction value of each hour on the day to be tested.

Images