Industry-specific load forecasting method and system based on type analysis and feature reconstruction

Abstract

The application belongs to the technical field of sub-industry load forecasting, and provides a sub-industry load forecasting method and system based on type analysis and feature reconstruction, which obtains power load data and meteorological feature data of each industry, and carries out pretreatment; classifies power load of different industries; calculates the maximum information coefficient between each meteorological feature variable and power load of different categories, selects features according to the value of the maximum information coefficient, and reconstructs a meteorological feature matrix; takes historical load data in a set time period of a day to be predicted as model input features, and combines them into the feature matrix; takes the combined feature matrix as input and load value as output, trains a deep learning model; selects a corresponding feature matrix according to the power type of a prediction target, and uses the trained deep learning model to perform day-ahead load forecasting. The application can realize day-ahead load forecasting of different industries.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of sub-industry load forecasting, and particularly relates to a sub-industry load forecasting method and system based on type analysis and feature reconstruction. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] Load forecasting is of great significance in the power system, which can help power operators accurately predict future load demand, thereby optimizing power production, dispatching and distribution. Through effective load forecasting, the rational allocation of power resources can be realized, the stability and reliability of the power grid can be improved, and the operating cost can be reduced.

[0004] However, the electricity demand of different industries is significantly different, and it is difficult to accurately match the actual electricity consumption of each industry, which may lead to excessive or insufficient electricity, thereby affecting the production efficiency and cost control of enterprises. Therefore, it is particularly important to predict the electricity load of sub-industry. Through accurate prediction of sub-industry electricity load, rational allocation of power resources, optimization of power purchase plan, effective avoidance of risks caused by price fluctuations, improvement of power purchase efficiency, and accurate satisfaction of industry electricity demand can be achieved.

[0005] Artificial intelligence algorithms based on machine learning and deep learning are the main prediction methods at present. However, the current artificial intelligence model usually directly takes the historical load sequence and its related meteorological data sequence as input features when training the model, and usually assumes that the influence degree of the same input feature at different time points is the same, which limits the model's ability to identify load fluctuations caused by meteorological factors to some extent. In addition, the electricity characteristics of different industries are significantly different, and the electricity demand of each industry is greatly affected by production process, business cycle, equipment load and other factors, resulting in different volatility and stability of load characteristics.

[0006] Therefore, it is often difficult to achieve ideal results by using the same model parameters to predict the electricity load of each industry. Therefore, how to accurately predict the electricity load according to the electricity characteristics of different industries has become a problem to be solved in the prior art. SUMMARY

[0007] In order to solve the above problems, the present application proposes a sub-industry load forecasting method and system based on type analysis and feature reconstruction, which can realize day-ahead load forecasting of different industries and provide reference information for power dispatching and production planning, as well as power purchase transactions.

[0008] According to some embodiments, the present application adopts the following technical solutions:

[0009] A method for industry-specific load forecasting based on type analysis and feature reconstruction, comprising the following steps:

[0010] Obtaining the power load data and meteorological feature data of each industry and performing normalization processing;

[0011] Classifying the power load of different industries;

[0012] Calculating the maximum information coefficient between each meteorological feature variable and different category loads, selecting features according to the value of the maximum information coefficient, and reconstructing the meteorological feature matrix;

[0013] Taking the historical load data within a set time period from the day to be predicted as the model input features and merging them into the feature matrix;

[0014] Taking the merged feature matrix as the input and the load value as the output, training the deep learning model;

[0015] According to the power consumption type of the prediction target, selecting the corresponding feature matrix and using the trained deep learning model to obtain the day-ahead prediction result of the load.

[0016] As an optional implementation, the process of classifying the loads of different industries includes dividing the loads of different industries into the following categories according to load characteristics:

[0017] Double-peak type: having two power consumption peaks exceeding a set value within a predetermined time;

[0018] Continuous power consumption type: the ratio of the average load in a continuous period to the average load throughout the day exceeds a set value;

[0019] Midday power consumption type: the ratio of the average load in the midday of a set time to the average load throughout the day exceeds a set threshold.

[0020] As an optional implementation, the process of calculating the maximum information coefficient between each meteorological feature variable and different category loads includes taking the load value y as the target variable, x and the meteorological feature variable to be screened as the input variable, and calculating the maximum information coefficient between each meteorological feature variable and the load data x to reflect the influence of the input variable y on the target variable.

[0021] As an optional implementation, the process of reconstructing the meteorological feature matrix includes:

[0022] Constructing the load sequence and meteorological feature sequence under different power consumption types;

[0023] Select meteorological characteristic variables, choosing those with the maximum information coefficient greater than a predetermined value, and determine the initial meteorological characteristic matrix.

[0024] The maximum information coefficient of different meteorological characteristic variables at each sampling point is used as the weight to measure the degree of influence of each meteorological characteristic, and a weight matrix is ​​constructed.

[0025] The final meteorological feature matrix is ​​obtained by multiplying the corresponding elements of the initial meteorological feature matrix and the weight matrix.

[0026] As a further step, in the process of selecting meteorological feature variables whose maximum information coefficient value is greater than a predetermined value, the maximum information coefficient value between different meteorological features and load data is obtained by calculating the data of multiple sampling points within a set time period. Then, the average maximum information coefficient value of each meteorological feature within the set time period is calculated, and different features are sorted according to the average maximum information coefficient value. Feature variables whose average maximum information coefficient value is greater than the predetermined value are selected to form the initial meteorological feature matrix of the prediction model.

[0027] As an alternative implementation, the deep learning model is a bidirectional long short-term memory network model. During the training process of the bidirectional long short-term memory network model, the memory units and hidden layers are updated each time the bidirectional long short-term memory network model is trained, using forget gates, input gates and output gates.

[0028] Furthermore, the calculation formula for the bidirectional long short-term memory network model is as follows:

[0029] ;

[0030] ;

[0031] ;

[0032] in, and They are respectively in t Load prediction values ​​for forward and backward long short-term memory networks at each time step; for Load forecast at any given time; and These represent the forward and backward computation processes of the Long Short-Term Memory (LSTM) network, respectively. for t Input characteristics at any given moment; and They are respectively t Output weights for both forward and backward iterations at each time step; for t A unit of memory for a specific moment; for tThe load prediction value output by the time bidirectional long short-term memory network model finally.

[0033] An industry-specific load prediction system based on type analysis and feature reconstruction comprises:

[0034] A data acquisition module is configured to acquire load data and meteorological feature data of various industries and perform normalization processing.

[0035] A type analysis module is configured to classify loads of different industries.

[0036] A feature reconstruction module is configured to calculate the maximum information coefficient between each meteorological feature variable and different category loads, select features according to the value of the maximum information coefficient, and reconstruct a meteorological feature matrix.

[0037] A data merging module is configured to merge historical load data within a set time period from a day to be predicted into a feature matrix as model input features.

[0038] A prediction model training module is configured to train a deep learning model with the merged feature matrix as input and load values as output.

[0039] A day-ahead prediction module is configured to select a corresponding feature matrix according to the power consumption type of a prediction target and obtain a day-ahead prediction result of the load by using the trained deep learning model.

[0040] A computer readable storage medium for storing computer instructions, which, when executed by a processor, completes the steps in the above method.

[0041] An electronic device comprising a memory and a processor, and computer instructions stored on the memory and running on the processor, which, when executed by the processor, complete the steps in the above method.

[0042] Compared with the prior art, the beneficial effects of the present application are:

[0043] The present application identifies different typical power consumption types by analyzing the power consumption characteristics of different industries and formulates corresponding discrimination conditions to realize the classification of industry loads.

[0044] To further improve the prediction accuracy, the application adopts a bidirectional long short-term memory network (BI-LSTM), which fully utilizes its bidirectional information processing capability and effectively captures the nonlinear relationship between the load and the input features. Experimental results show that the BI-LSTM model after feature reconstruction exhibits excellent prediction performance under different power consumption types, which can provide reference information for power dispatching and production planning, and power purchase transactions.

[0045] In order to make the above-mentioned objects, features and advantages of the present application more apparent, a preferred embodiment is described below, and the accompanying drawings are described in detail as follows. BRIEF DESCRIPTION OF DRAWINGS

[0046] The drawings accompanying the specification of this application form a part thereof and serve to further understand the present application, the illustrative embodiments thereof and the description thereof, and do not constitute an improper limitation of the present application.

[0047] Figure 1 It is a schematic diagram of the first form of bimodal power consumption characteristics in an embodiment;

[0048] Figure 2 It is a schematic diagram of the second form of bimodal power consumption characteristics in an embodiment;

[0049] Figure 3 It is a schematic diagram of continuous power consumption characteristics in an embodiment;

[0050] Figure 4 It is a schematic diagram of midday power consumption characteristics in an embodiment;

[0051] Figure 5 It is a prediction flowchart of load prediction in an embodiment;

[0052] Figure 6 It is a schematic diagram of the maximum information coefficient calculation result of the first bimodal power consumption type in an embodiment;

[0053] Figure 7 It is a schematic diagram of the maximum information coefficient calculation result of the second bimodal power consumption type in an embodiment;

[0054] Figure 8 It is a schematic diagram of the maximum information coefficient calculation result of the continuous power consumption type in an embodiment;

[0055] Figure 9 It is a schematic diagram of the maximum information coefficient calculation result of the midday power consumption type in an embodiment;

[0056] Figure 10 It is a schematic diagram of the load prediction result of the first bimodal power consumption type in an embodiment;

[0057] Figure 11 Fig. 2 is a schematic diagram of a load prediction result in a second type of bimodal power consumption type in an embodiment;

[0058] Figure 12 Fig. 3 is a schematic diagram of a load prediction result in a continuous power consumption type in an embodiment;

[0059] Figure 13 Fig. 4 is a schematic diagram of a load prediction result in a midday power consumption type in an embodiment. DETAILED DESCRIPTION

[0060] The application will be further described below in conjunction with the accompanying drawings and embodiments.

[0061] It should be noted that the following detailed description is illustrative only and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0062] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments in accordance with the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, steps, operations, elements, components, and / or groups thereof, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof.

[0063] The embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0064] Embodiment One

[0065] Different industries exhibit unique power consumption characteristics due to their different operating modes and demand characteristics. These characteristics not only affect the energy costs of enterprises, but also have an important impact on the electricity market and grid management. In order to more effectively manage and predict power demand, it is particularly important to divide different power consumption types according to power consumption characteristics and develop discrimination standards.

[0066] There are differences in power consumption levels under different industries, and different meteorological element units. In order to reduce the influence of the above dimension and unit on load prediction, first, the load data of each industry and the meteorological element data are normalized, as shown in formula (1):

[0067] (1)

[0068] In the formula, is the normalized value of each load or meteorological data, is the original data thereof, and Max and Min represent the maximum and minimum values of the original data of the corresponding load or weather sequence, respectively.

[0069] Due to the variety of different industries, in order to reduce the modeling workload, it is necessary to classify them. According to different load characteristics, the following typical power consumption types can be classified:

[0070] (1) Double-peak type

[0071] This type of power consumption pattern shows two distinct peaks. The first form of two peaks occurs at 9 to 11 am and 14 to 16 pm, and this double-peak phenomenon is usually consistent with the work and rest patterns of weekdays, as shown in Figure 1 , where each sample is a power consumer of this type. The second form of two peaks occurs at 10 to 14 pm and 18 to 22 pm, and this power consumption characteristic is usually related to the catering and accommodation industry, as shown in Figure 2 , where each sample is a power consumer of this type.

[0072] The double-peak type can be distinguished by comparing the two peak period loads with the daily average load, and the distinguishing criterion is defined by equation (2):

[0073] (2)

[0074] In the equation, and are the average loads of the two peak periods, is the average load for the whole day, is a set proportion threshold to identify two distinct peaks.

[0075] (2) Continuous power consumption type

[0076] The main feature of this type of power consumption pattern is that the load usage remains at a high level from the afternoon to the evening, as shown in Figure 3 , where each sample is a power consumer of this type. This characteristic is usually related to continuous business activities or production operations, reflecting the high power demand of some industries during this period.

[0077] The continuous power consumption type can be distinguished by the proportion of the average load in the continuous period to the average load for the whole day, and the distinguishing criterion is defined by equation (3):

[0078] (3)

[0079] In the equation, is the average load of the continuous period from the afternoon to the evening, is a set proportion threshold, which should be set high enough to ensure the distinction of the true continuous peak power consumption characteristics.

[0080] (3) Lunchtime type

[0081] The characteristic of this type of power consumption is that the power load increases sharply during lunch time, i.e. between 11:00 and 14:00, as shown in FIG. 2. The power consumption peak in this period is mainly due to the centralized start-up of power equipment, such as kitchen equipment and air conditioning systems. Figure 4

[0082] The lunchtime type can be distinguished by the ratio of the lunchtime average load to the daily average load, which is defined by formula (4):

[0083] (4)

[0084] In the formula, is the average load between 11:00 and 14:00, is a set proportion threshold.

[0085] Due to the different operation modes of different industries, the meteorological influence factors under different power consumption modes have significant differences. Therefore, the maximum information coefficient method is used to reconstruct the meteorological characteristics in this embodiment. By calculating the maximum information coefficient of different meteorological factors and the power load under different power consumption characteristics, the influence degree of meteorological factor characteristics on the power load is quantified, and the meteorological characteristics are sorted according to the influence degree. A group of characteristic variables with the most significant influence on the power load is selected as the meteorological input feature matrix.

[0086] The maximal information coefficient (MIC) is a correlation measure index that can widely mine the multi-dimensional dependence relationship between variables in large data sets. Its definition is as follows: let be a finite two-dimensional data set, x and y be the intervals divided in the current two-dimensional space X, Y direction; G is the dimensional grid formed by this; is the set formed by different grid division methods, then define:

[0087] (5)

[0088] In the formula, represents the distribution of the data set D on the divided grid G ; represents the mutual information of . Let the maximum normalized mutual information matrix of the data set D under different division intervals be ​, whose elements are

[0089] (6)

[0090] wherein, denotes the smaller one of the two. Then the maximum information coefficient can be defined as:

[0091] (7)

[0092] wherein, n is the number of samples of the data set; is the upper limit value of the grid division, which is generally in the range of , As the value of increases, the accuracy of the MIC in measuring the correlation increases, but the calculation complexity also increases. In this embodiment, is taken, which can balance the accuracy and calculation complexity, thereby achieving good results.

[0093] When using the MIC as the correlation measurement standard for the selection of the load day-ahead prediction weather features, the load value y is taken as the target variable, x is the prediction model weather input variable to be screened, and the maximum information coefficient between each variable and the load data is calculated to reflect the influence degree of the input variable x on the target variable y , thereby realizing the input feature selection.

[0094] The MIC overcomes the limitation that the Pearson correlation coefficient cannot reveal the nonlinear relationship, and can more effectively extract valuable input features when processing the load data under the influence of multi-dimensional weather features. After completing the feature selection, the MIC value is taken as the weight of each weather input variable, and the weather input variable is reconstructed according to the weight. The reconstructed weather input matrix contains each weather feature variable and its corresponding weight, thereby reflecting the influence degree of different variables on the load in the prediction model. The specific steps of the feature selection and reconstruction in this embodiment are as follows:

[0095] (1) Constructing the load sequence and the weather feature sequence under different types of electricity consumption.

[0096] For the kth type of electricity consumption, the load value of the mth sampling point of each day in the historical sample is extracted, sorted in time sequence, and the sampling point level load sequence of the nth time is obtained: i i

[0097] (8) ​​

[0098] In the formula, This represents the k-th type of electricity consumption characteristic on day n. ( Normalized load values ​​at ) sampling points.

[0099] For each meteorological characteristic data affecting the load, the same method is used to reconstruct it. For the first... m Extract all date sampling points from the feature sequence. The value can be obtained as follows:

[0100] (9)

[0101] In the formula, Representing the m Type of feature in the first n Heavenly Normalized data from each sampling point.

[0102] (2) Select meteorological features and determine the initial meteorological feature dataset.

[0103] For the k Electricity consumption characteristics, with For the target sequence, For the candidate feature sequences, calculate and The MIC value was calculated by analyzing data from 24 sampling points throughout the day to obtain the MIC values ​​between different meteorological characteristics and load data. Then, the average MIC value of each meteorological characteristic over 24 hours was calculated, and the characteristics were sorted according to their average value. The characteristic variables with an average MIC greater than 0.3 were selected to form the initial meteorological input feature set for the prediction model, thus obtaining the initial meteorological input feature set corresponding to the daily 24-hour load. .

[0104] (10)

[0105] In the formula, Representing the k Under the characteristics of electricity consumption, the first q The meteorological characteristics on the day i Normalized values ​​of each sampling point.

[0106] (3) Reconstruct input features based on MIC values.

[0107] Using the MIC values ​​of different variables at each sampling point as weights to measure the influence of each meteorological feature, an input weight matrix is ​​constructed. The MIC values ​​of the different variables at each sampling point are then used as weights to measure the degree of influence of each meteorological feature. q The input features and load in the first... i The MIC values ​​of each sampling point are used as the weight values ​​of the input feature at that time, forming the weight matrix of the input feature. M. Defined as:

[0108] (11)

[0109] Thus, the weights of all variables constitute the weight matrix under the kth type of power consumption .

[0110] (12)

[0111] (4) Determine the input feature data set.

[0112] Take the feature matrix in step three and step four and the feature weight matrix , multiply the corresponding positions of each element of the two matrices to obtain the final weather input matrix , that is

[0113] (13)

[0114] In the formula, * represents the Hadamard product of two matrices.

[0115] Since the power consumption of the same industry on the adjacent date is similar, the load data between adjacent dates has great correlation, therefore, when performing load prediction, the historical load data of the same industry of the nearest three days of the same type (both working days or both non-working days) of the day to be predicted is also used as model input features, and is merged into the feature matrix.

[0116] After determining the input features, the prediction model is trained, and in this embodiment, a bidirectional long short-term memory (BILSTM) is selected as the prediction model.

[0117] LSTM is a commonly used time series prediction model, which can effectively capture the nonlinear change relationship between the target variable and all input features from historical data. According to the load characteristics of different industries, the current load value is related to the information of historical time and future time, so it is suitable for LSTM model. Through its unique "gated unit" structure, LSTM can flexibly add or delete information, thereby realizing the training with memory function. This feature makes LSTM more accurate in learning long-term dependencies, and can also more effectively capture nonlinear dependencies. In the training process of LSTM, each training not only depends on the hidden layer of the last training, but also introduces a memory unit . Through the three gate units of the forgetting gate, the input gate and the output gate, LSTM can effectively update the memory unit and the hidden layer This enables the dynamic transmission and storage of information.

[0118] Building upon the LSTM model, this embodiment further selects BILSTM, which can simultaneously consider bidirectional time information, as the day-ahead load forecasting model. The input data of BI-LSTM is processed through forward and backward LSTM networks respectively, and the final output contains the concatenation of information from both directions, thereby improving prediction accuracy. BI-LSTM is a special recurrent neural network structure that, by combining forward and backward LSTM networks, can simultaneously capture the forward and backward time dependencies in the sequence. By processing past (forward) and future (backward) information in parallel, BI-LSTM enhances the modeling capability of sequence data, thus demonstrating superior performance in sequence modeling tasks.

[0119] The formula for calculating BILSTM is shown below.

[0120] (14)

[0121] (15)

[0122] (16)

[0123] In the formula, and They are respectively in t Load forecast values ​​of forward and backward LSTM cells at any given time; for Load forecast at any given time; and These represent the forward and backward computation processes of the LSTM, respectively. for t Input characteristics at any given moment; and They are respectively t Output weights for both forward and backward iterations at each time step; for t A unit of memory for a specific moment; for t The load prediction value finally output by the BILSTM unit at that time.

[0124] according to Figure 5 The process shown implements day-ahead load forecasting, and the forecast results are evaluated using normalized mean absolute error (NMAE) and normalized root mean square error (NRMSE), as shown in equations (17) and (18).

[0125] (17)

[0126] (18)

[0127] In the formula, and for i Real and predicted loads normalized to time; N The total number of samples.

[0128] Example 2

[0129] This embodiment uses industry-specific load data and corresponding meteorological data from a certain city from 2020 to 2022 to conduct a day-ahead load forecasting case study verification. The load data resolution is 1 hour, and the meteorological data resolution is the same as the load data, including load influence characteristics such as temperature, humidity, wind speed, and radiation.

[0130] First, a candidate feature set is constructed based on existing meteorological data. Feature selection is performed, and the data is reconstructed into new input features according to weights. The candidate feature set contains 14 different variables. The minimum input characteristic (MIC) values ​​between different meteorological features and load data are calculated for various electricity consumption types. Figure 6-9 As shown.

[0131] by The meteorological input features for the prediction model are selected based on the threshold. When the MIC is less than 0.3, the meteorological feature is considered to have a small impact on the load and can be ignored. The results of the selection of meteorological features under different electricity consumption types are shown in Table 1.

[0132] Table 1 Feature selection results

[0133]

[0134] It can be seen that the MIC values ​​of various meteorological impact characteristics differ under different types of electricity consumption, and the selected high-impact characteristics also differ, reflecting the differences in electricity consumption behavior in different industries.

[0135] Since the electricity load characteristics of adjacent dates are similar, historical load data from the three days closest to the forecast date (both weekdays or both weekend days) of the same type are also used as input features. The MIC value is used as the weight to reconstruct the input features, which are then used as the input to the final model. For different electricity consumption types, the selected meteorological features are combined with historical load information to reconstruct the input feature matrix.

[0136] After confirming the input features, a Bi-LSTM network was used to conduct day-ahead load prediction experiments for various electricity consumption types. For each electricity consumption type, training, testing, and validation sets were constructed. The Bi-LSTM model was denoted as M1, and the LSTM model was used as the control model M2. Simultaneously, the Bi-LSTM model without feature reconstruction was used as M3. The day-ahead load prediction results and errors for various electricity consumption types are compared below. Figure 10-13and shown in Table 2.

[0137] Table 2 prediction error comparison

[0138]

[0139] It can be seen that under different power consumption types, the proposed model can achieve good prediction effect and smaller prediction error. Compared with M3 without using the feature selection method proposed in the embodiment, the NMAE and NRMSE are reduced by 3.009%-6.880% and 3.392%-8.624% respectively, which proves the effectiveness of the MIC feature reconstruction method proposed in the embodiment.

[0140] In the bimodal prediction, the model proposed in the embodiment can accurately capture the peak and valley values of the morning and evening peak periods, because LSTM can effectively learn and remember long-term time series dependencies, so that the model can not only predict the magnitude of the load, but also accurately predict the time point of the load rising and falling; in the continuous power consumption type prediction, the model proposed in the embodiment shows good recognition ability to the continuous high load state, and can maintain accurate tracking of the load trend without significant prediction deviation, which is of great significance for energy management and cost control of continuously operating commercial facilities or industrial production; in the midday power consumption type prediction, the model proposed in the embodiment can accurately capture the sharp rise and subsequent decline of the load at noon, and shows rapid response ability to the midday power consumption peak.

[0141] Embodiment three

[0142] The industry-specific load prediction system based on type analysis and feature reconstruction comprises:

[0143] The data acquisition module is configured to acquire power consumption load data and meteorological feature data of each industry and perform normalization processing;

[0144] The type analysis module is configured to classify the power consumption load of different industries;

[0145] The feature reconstruction module is configured to calculate the maximum information coefficient between each meteorological feature variable and different category loads, select features according to the value of the maximum information coefficient, and reconstruct the meteorological feature matrix;

[0146] The data merging module is configured to merge the historical load data within a set time period from the day to be predicted into the feature matrix as model input features;

[0147] The prediction model training module is configured to train the deep learning model by taking the merged feature matrix as input and the load value as output;

[0148] The day-ahead prediction module is configured to select a corresponding feature matrix according to a power consumption type of the prediction target, and obtain a day-ahead prediction result of the power consumption load by using the trained deep learning model.

[0149] Those skilled in the art will understand that embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk memory, optical storage devices, etc.) containing computer-usable program code. CD - ROM

[0150] The present application is described in reference to flowcharts and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and combinations of flows and / or blocks in the flowcharts and / or block diagrams 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 apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce an apparatus that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in the flowcharts and / or block diagrams.

[0151] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction apparatus, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in the flowcharts and / or block diagrams.

[0152] These computer program instructions can also be loaded into a computer or other programmable data processing apparatus, so that a series of operation steps are performed on the computer or other programmable data processing apparatus to produce a computer-implemented process, so that the instructions executed on the computer or other programmable data processing apparatus provide steps for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in the flowcharts and / or block diagrams.

[0153] ​The above merely describes the preferred embodiments of the present application and is not used to limit the present application. Any modification, equivalent replacement, improvement, etc. made by those skilled in the art without departing from the spirit and principle of the present application shall fall within the protection scope of the present application.

Claims

1. A method for industry-specific load forecasting based on type analysis and feature reconstruction, characterized by, The method comprises the following steps: obtaining power load data and meteorological characteristic data of various industries and performing normalization processing; classifying power loads of different industries; the process of classifying power loads of different industries comprises the following steps: bimodal type: having two power peaks exceeding a set value within a predetermined time; continuous power type: the ratio of the average value of power load in a continuous period to the average value of full-day load exceeds a set value; midday power type: the ratio of the average value of load in a set time to the average value of full-day load exceeds a set threshold value; calculating the maximum information coefficient between each meteorological characteristic variable and power load of different categories, selecting features according to the value of the maximum information coefficient, and reconstructing a meteorological characteristic matrix; The process of calculating the maximum information coefficient between each meteorological feature variable and the load of different categories includes taking the load value y as the target variable, x the meteorological feature variable to be screened, calculating the maximum information coefficient between each meteorological feature variable and the load data to reflect the influence degree of the input variable x on the target variable y ; the process of reconstructing the meteorological characteristic matrix comprises the following steps: constructing load sequences and meteorological characteristic sequences under different power types; selecting meteorological characteristic variables, selecting meteorological characteristic variables with a maximum information coefficient value greater than a predetermined value to determine an initial meteorological characteristic matrix, in the process of selecting meteorological characteristic variables with a maximum information coefficient value greater than a predetermined value, the maximum information coefficient value between different meteorological characteristics and load data is obtained by calculating the data of multiple sampling points in a set time period, then the average maximum information coefficient value of each meteorological characteristic in the set time period is calculated, different features are sorted according to the average maximum information coefficient value, and the feature variables with an average maximum information coefficient value greater than a set value are selected to form the initial meteorological characteristic matrix of the prediction model; taking the maximum information coefficient value of each meteorological characteristic variable at each sampling point as the weight of the influence degree of each meteorological characteristic, and constructing a weight matrix; multiplying the corresponding position elements of the initial meteorological characteristic matrix and the weight matrix to obtain the final meteorological characteristic matrix; combining the historical load data in a set time period before the day to be predicted into the feature matrix as the model input feature; training a deep learning model by taking the combined feature matrix as the input and the load value as the output; selecting the corresponding feature matrix according to the power type of the prediction target, and obtaining the day-ahead prediction result of the power load by using the trained deep learning model; the deep learning model is a bidirectional long short-term memory network model, in the training process of the bidirectional long short-term memory network model, the bidirectional long short-term memory network model uses a forgetting gate, an input gate and an output gate to update a memory unit and a hidden layer each time; the calculation formula of the bidirectional long short-term memory network model is specifically as follows: ； ； ； wherein, and are the load prediction values of the forward and backward long short-term memory networks at time t ; is the load prediction value at time ; and represent the forward and backward calculation processes of the long short-term memory network, respectively; is the input feature at time t ; and are the forward and backward output weights at time t ; is the memory unit at time t ; is the load prediction value of the bidirectional long short-term memory network model at time t .

2. A load forecasting system for each industry based on type analysis and feature reconstruction, which adopts the load forecasting method for each industry based on type analysis and feature reconstruction as claimed in claim 1, characterized in that, comprising: a data acquisition module configured to obtain power load data and meteorological characteristic data of various industries and perform normalization processing; a type analysis module configured to classify power loads of different industries; the process of classifying power loads of different industries comprises the following steps: bimodal type: having two power peaks exceeding a set value within a predetermined time; continuous power type: the ratio of the average value of power load in a continuous period to the average value of full-day load exceeds a set value; Midday electricity type: the ratio of the average value of the midday load set time to the average value of the whole day load exceeds the set threshold value; The feature reconstruction module is configured to calculate the maximum information coefficient between each meteorological feature variable and different categories of electricity load, select features according to the value of the maximum information coefficient, and reconstruct the meteorological feature matrix; The process of calculating the maximum information coefficient between each meteorological feature variable and different categories of load includes taking the load value y as the target variable, x the meteorological feature variable to be screened, calculating the maximum information coefficient between each meteorological feature variable and the load data to reflect the degree of influence of the input variable x on the target variable y ; The process of reconstructing the meteorological feature matrix includes: Constructing the load sequence and the meteorological feature sequence under different electricity types; Selecting meteorological feature variables, selecting meteorological feature variables with a maximum information coefficient value greater than a predetermined value, and determining an initial meteorological feature matrix; In the process of selecting meteorological feature variables with a maximum information coefficient value greater than a predetermined value, the maximum information coefficient values between different meteorological features and load data are obtained by calculating the data of multiple sampling points in a set time period, then the average maximum information coefficient values of each meteorological feature in the set time period are calculated, and different features are sorted according to the average maximum information coefficient values, and the feature variables with an average maximum information coefficient value greater than a set value are selected to form an initial meteorological feature matrix of the prediction model; Taking the value of the maximum information coefficient of each meteorological feature variable at each sampling point as the weight of the influence degree of each meteorological feature, a weight matrix is constructed; Multiplying the corresponding position elements of the initial meteorological feature matrix and the weight matrix to obtain the final meteorological feature matrix; The data merging module is configured to merge the historical load data within a set time period from the day to be predicted into the feature matrix as model input features; The prediction model training module is configured to train the deep learning model by taking the merged feature matrix as input and the load value as output; The day-ahead prediction module is configured to select the corresponding feature matrix according to the electricity type of the prediction target, and obtain the day-ahead prediction result of the load by using the trained deep learning model.

3. A computer readable storage medium characterized by, A computer program product for storing computer instructions, which, when executed by a processor, completes the steps in the method of claim 1.

4. An electronic device, characterized by comprising: A computer program product for storing computer instructions, which, when executed by a processor, completes the steps in the method of claim 1.

Images