A method for short-term power forecasting based on generalized hierarchical multi-granularity

Abstract

The invention discloses a method for short-term power prediction based on generalized hierarchical multi-granularity, which comprises the following steps: obtaining power data including factors affecting power load and power load values; screening information affecting power load according to attribute importance and similarity measurement; establishing concept hierarchical trees for factors affecting power load and power load values ​​respectively; constructing attribute combination granules affecting power load according to time series relationship; selecting the coarsest granularity level with unchanged attribute importance affecting power load according to the reverse order of hierarchical importance sorting; mining rules meeting the requirements and making predictions by using a multi-granularity method, the invention can greatly reduce the calculation time and the amount of data processed by using the structural characteristics of the hierarchical model; mining rules different from flat data tables by using the multi-perspective characteristics of multi-granularity, and discovering more meaningful information.

Description

Technical Field

[0001] The invention relates to a method for short-term power forecasting based on generalized hierarchical multi-granularity, in particular to a method for short-term power forecasting based on generalized hierarchical multi-granularity. Background Art

[0002] Existing short-term power load forecasting methods are based on historical data, and model building is based on the most fine-grained data. Decision rule mining is an important purpose of building a short-term power forecasting model. We need to mine useful and valuable rules from a large amount of even chaotic data, which places great demands on existing hardware and model building. The factors that affect power load have many attributes, and also generate a lot of redundant information. Filtering redundant information is very important for model building. For large-scale data, power load rules are generally mined from flat data tables, but many times the mined rules are not conducive to understanding and the rules are too simplistic. Summary of the invention

[0003] The purpose of the present invention is to improve and innovate the shortcomings and problems existing in the background technology, and to provide a short-term power forecasting method based on generalized hierarchical multi-granularity.

[0004] To achieve the above object, the present invention provides the following technical solution: a method for short-term power forecasting based on generalized hierarchical multi-granularity, the method specifically comprising the following steps:

[0005] Step S1, preprocessing the power load data and discretizing the data;

[0006] Step S2, according to the importance of the attribute affecting the power load and the distance of the granular structure affecting the power load, select the attribute affecting the power load whose importance is greater than 0.2;

[0007] Step S3, selecting the attributes that affect the power load after screening, and establishing a hierarchical model of the power load through a concept hierarchy tree;

[0008] Step S4: According to user needs and questions, first select the level of decision attribute power load, then sort the level attributes according to the importance of the level attributes affecting the power load, select the condition attribute level, and construct the time series condition granule according to the time series;

[0009] Step S5: Use a multi-granularity approach and apply different strategies to mine rules.

[0010] A further solution is that the data preprocessing in step S1, discretization of the data specifically includes:

[0011] The attribute values ​​that may affect the power load (climate, holidays, electricity price policy, emergencies, power load) are divided into discrete values ​​according to certain standards, represented by 0, 1,..., n. For example, the fog level can be divided into light fog, fog, heavy fog, dense fog, and very dense fog; the date type can be divided into weekdays and weekends; the power load can be divided into extremely low, relatively low, normal, relatively high, and extremely high. According to needs, the power load value at a certain moment or the daily power load can be used as a decision attribute; other attributes are discretized accordingly.

[0012] A further solution is that the step S2 of screening the attributes affecting the power load specifically includes:

[0013] Step 21: Define the power load D and the attribute a that affects the power load i Dependencies:

[0014] The power load D affects the power load attribute a i The dependency definition is as follows:

[0015]

[0016] Among them, |POS a (D) | represents the attribute a that affects the power load i The number of positive domains of the relative decision attribute power load D, |U| represents the number of selected power data;

[0017] Step 22: Calculate the importance of attributes that affect the power load:

[0018] Define the importance of attributes that affect the power load;

[0019] SGF(a i , C, D) = R(C, D) - R(C - {a i}, D).;

[0020] Where SGF(a, C, D) represents attribute a i In the conditional attribute C = {a 1 , a 2 , ..., a n} relative to the importance of the decision attribute power load D;

[0021] Step 23: Delete the attributes that affect the power load and whose importance is less than or equal to 0.2:

[0022] Sort by importance and delete the attributes that affect the power load if the attribute importance is less than or equal to 0.2;

[0023] Step 24: Calculate the distance between the attribute combinations that affect the power load:

[0024] Calculate the particle distance that affects the power load:

[0025]

[0026] in, Represents the distance between different particles, K i Indicates the attribute a that affects the power load i Derived particle structure U / a i = {G i (x 1 ), G i (x 2 ), .., G i (x l ), ..., G i (x max )}.

[0027] Step S25: Calculate the similarity of attributes affecting the power load:

[0028] Based on the distance between different particles, the similarity of two attributes is calculated as follows:

[0029] Among them, Sim(K i , K j ) represents the particle structure K i and K j The similarity of d(K i , K j ) represents the particle structure K i and K j distance.

[0030] Step S26: Filter the attributes that have an impact on the power load and whose importance is greater than 0.2. Select the attribute with the smallest information entropy value.

[0031] A further solution is that the step S3 of establishing the hierarchical model specifically includes:

[0032] Select the filtered attributes that affect the power load;

[0033] Establish a conceptual hierarchy tree of power load:

[0034] Each attribute affecting the power load and the power load decision attribute are hierarchically clustered to establish a conceptual hierarchical tree.

[0035] A further solution is that the level selection in step S4 specifically includes:

[0036] Step S41: The importance of the attributes affecting the power load at the kth layer in the hierarchical model of the power load is as follows:

[0037]

[0038] Among them, SGF k (a i , C, D) represent the importance of attributes affecting power load in the hierarchical model, Indicates a i Take the attributes of the kth layer, and the other attributes take the most fine-grained attributes. Indicates the attribute a that will affect the power load i The remaining conditional attributes removed;

[0039] Step S42: The importance of attributes affecting the power load in the hierarchical model of the power load is as follows:

[0040]

[0041] Among them, SGF (a i , C, D) represent the hierarchical attributes a that affect the power load i Importance, w j Represents the weight of the jth level of the attribute that affects the power load, and the sum of the weights of each level Here each level is given equal weight Then sort;

[0042] Step S43: construct the attribute combination granules that affect the power load in the time series (day): generalize each attribute level to the coarsest granularity level with unchanged importance in descending order of the attribute importance that affects the power load, and construct the combination granules that affect the power load in the time series T = {1, 2, ..., t}, respectively, expressed as:

[0043] CG={cg 1 , cg 2 , ..., cgt}

[0044] The time sequence T = {1, 2, ..., t} represents the sequence of the previous t moments.

[0045] A further solution is that the multi-granularity method used in step S5 specifically includes:

[0046] A multi-granularity approach is adopted to match rules according to different granularities.

[0047]

[0048]

[0049] in Respectively represent the multi-granularity methods that adopt optimistic and pessimistic strategies, Represented by the conditional particle cg i The derived equivalence class, Represents the equivalence class derived from the j-th value of the h-th layer of the decision attribute power load.

[0050] A further solution is that the attribute combination particles of different influencing power load are in a parallel relationship, and there is no inclusion relationship between the particles.

[0051] A further solution is that the decision attribute level selection selects the level of power load D according to different user needs. The coarse-grained level is suitable for user understanding, and the fine-grained level is more suitable for detailed decision making.

[0052] A further solution is that the importance does not change, and for the attribute, the levels are continuously generalized, the importance is calculated until the level h where the importance changes, and then the h-1 level is selected as the final generalized level.

[0053] A further solution is that, in the multi-granularity method, the optimistic strategy is a more relaxed condition, which must provide all possible situations based on historical data. We can select the maximum or minimum value of all possibilities according to actual needs; the pessimistic strategy is a more stringent strategy, which matches historical data and mines more accurate rules; the multi-granularity method selects the t value of the time series T={1, 2, ..., t}, t>0 as the maximum value of the minimum rules obtained by the pessimistic strategy.

[0054] The beneficial effects of the present invention are as follows: the existing data processing is based on flat data tables. The rules mined from flat data tables are not easy to understand and it is difficult to reflect the hierarchical characteristics of the data. Generally speaking, the decision rules mined from the hierarchical model can analyze the power load from different levels and different perspectives. The extension of multi-granularity rough sets on the classic single-granularity rough sets is defined by using multiple equivalent relations on the domain. These equivalent relations can be selected according to the needs or problems of the user. Therefore, we convert the flat data table into a multidimensional data cube to better discover valuable information. In view of the limitations of existing data tables, such as large data volume, flatness, simplification, and missing values, the generalized hierarchical multi-granularity short-term power forecasting method of the present invention utilizes the structural characteristics of the hierarchical model to greatly reduce the calculation time and the amount of data processed; and utilizes the multi-granularity multi-perspective characteristics to mine out rules that are different from flat data tables and discover more meaningful information. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a business flow chart of the power short-term forecasting method based on generalized hierarchical multi-granularity of the present invention;

[0056] Figure 2It is a schematic diagram of the hierarchical model structure of the present invention;

[0057] Figure 3 It is a schematic diagram of the multi-level and multi-view model of the present invention;

[0058] Figure 4 It is a conceptual hierarchy tree established by the present invention. DETAILED DESCRIPTION

[0059] The present invention will be further explained below in conjunction with the accompanying drawings and embodiments.

[0060] Figure 1 This is the business process of the present embodiment based on the generalized hierarchical and multi-granular short-term electricity forecasting method.

[0061] This embodiment first screens the information, calls the factors that affect the power load as conditional attributes, and the power load as decision attributes, respectively, and establishes a concept hierarchy tree, and then replaces the low-level concepts with high-level concepts to reduce the size of the data table; then selects the corresponding conditional attribute and decision attribute hierarchy to construct the time series condition granules; finally, uses the multi-granularity method to establish the connection between the time series condition granules to perform short-term load forecasting;

[0062] A method for short-term power forecasting based on generalized hierarchical multi-granularity specifically includes the following steps:

[0063] Step S1, preprocessing the power load data and discretizing the data;

[0064] Specifically, the factors that affect the power load (climate, holidays, electricity price policies, emergencies, power load) are divided into discrete values ​​according to certain standards, represented by 0, 1,..., n. For example, the fog level can be divided into light fog, fog, heavy fog, dense fog, and very dense fog; the date type can be divided into weekdays and weekends; the power load can be divided into extremely low, relatively low, normal, relatively high, and extremely high. According to needs, the power load value at a certain moment or the daily power load can be used as a decision attribute; other attributes are discretized accordingly.

[0065] Step S2, according to the importance of the attribute affecting the power load and the distance of the granular structure of the attribute affecting the power load, select the attribute affecting the power load whose importance is greater than 0.2;

[0066] Step S3, selecting the attributes that affect the power load after screening, and establishing a hierarchical model of the power load through a concept hierarchy tree;

[0067] Specifically, firstly, the filtered attributes affecting the power load are selected;

[0068] Then, a conceptual hierarchical tree of power load is established: each attribute affecting power load and power load decision attribute is hierarchically clustered to establish a conceptual hierarchical tree.

[0069] Step S4: According to user needs and questions, first select the level of decision attribute power load, then sort the level attributes according to their importance, select the condition attribute level, and construct the time series condition granule according to the time series;

[0070] Step S5: Use a multi-granularity approach and apply different strategies to mine rules.

[0071] In specific implementation, the step S2 of screening the attributes affecting the power load can be specifically implemented according to the following sub-steps S21-526:

[0072] Step 21: Define the power load D and the attribute a that affects the power load i Dependencies:

[0073] The power load D affects the power load attribute a i The dependency definition is as follows:

[0074]

[0075] Among them, |POS a (D) | represents the attribute a that affects the power load i The number of positive domains of the relative decision attribute power load D, |U| represents the number of selected power data;

[0076] Step 22: Calculate the importance of attributes that affect the power load:

[0077] Define the importance of attributes that affect the power load;

[0078] SGF(a i , C, D) = R(C, D) - R(c - {a i}, D).:

[0079] Where SGF(a, C, D) represents attribute a i In the conditional attribute C = {a 1 , a 2 , ..., a n} relative to the importance of the decision attribute power load D;

[0080] Step 23: Delete the attributes that affect the power load and whose importance is less than or equal to 0.2:

[0081] Sort by importance and delete the attributes that affect the power load if the attribute importance is less than or equal to 0.2;

[0082] Step 24: Calculate the distance between attribute particles that affect the power load:

[0083] Calculate the particle distance that affects the power load:

[0084]

[0085] in, Represents the distance between different particles, K i Indicates the attribute a that affects the power load i Derived particle structure U / a i = {G i (x 1 ), G i (x 2 ), ..., G i (x l ), ..., G i (x max )};

[0086] Step S25: Calculate the similarity of attributes affecting the power load:

[0087] Based on the distance between different particles, the similarity of two attributes is calculated as follows:

[0088] Sim(K i , K j )=1-d(K i , K j )

[0089] Among them, Sim(K i , K j ) represents the particle structure K i and K j The similarity of d(K i , K j ) represents the particle structure K i and K j distance.

[0090] Step S26: Filter the attribute with the smallest value of information entropy affecting the power load:

[0091] when Select the attribute with the smallest information entropy value.

[0092] In specific implementation, the level selection in step S4 can be specifically implemented according to the following sub-steps S41-S43:

[0093] Step S41: The importance of the attributes affecting the power load at the kth layer in the hierarchical model of the power load is as follows:

[0094]

[0095] Among them, SGF k (a i , C, D) represent the importance of attributes affecting power load in the hierarchical model, Indicates a i Take the attributes of the kth layer, and take the most fine-grained attributes for other attributes. Indicates the attribute a that will affect the power load i The remaining conditional attributes are removed;

[0096] Step S42: The importance of attributes affecting the power load in the hierarchical model of the power load is as follows:

[0097]

[0098] Among them, SGF (a i , C, D) represents the hierarchical attribute a i Importance, w j Represents the weight of the attribute at the jth level, and the sum of the weights of each level Here each level is given equal weight Then sort;

[0099] Step S43: construct the attribute combination particle that affects the power load in the time series (day):

[0100] In descending order of importance affecting the power load, each attribute level is generalized to the coarsest granularity level where the importance does not change, and conditional granules with time series T = {1, 2, ..., t} are constructed respectively, which can be expressed as:

[0101] CG={cg 1 , cg 2 , ..., cg t}

[0102] The time sequence T = {1, 2, ..., t} represents the sequence of the previous t moments.

[0103] In specific implementation, the use of the multi-granularity method in step S5 can be specifically implemented according to the following sub-steps S51-S43:

[0104] In specific implementation, the level selection in step S4 can be specifically implemented according to the following steps:

[0105] Adopting a multi-granularity approach, matching rules according to different granules;

[0106]

[0107]

[0108] in Respectively represent the multi-granularity method of taking optimistic and pessimistic strategies, Represented by the conditional particle cg i The derived equivalence class, Represents the equivalence class derived from the j-th value of the h-th layer of the decision attribute power load.

[0109] It should be pointed out that the different granularities are in a parallel relationship, and there is no inclusion relationship between the granularities.

[0110] It should be pointed out that the decision attribute level selection selects the level of power load D according to different user needs. The coarse-grained level is suitable for user understanding, and the fine-grained level is more suitable for detailed decision-making.

[0111] It should be pointed out that the importance does not change. For the attribute, the level is continuously generalized and the importance is calculated until the level h where the importance changes, and then the h-1 level is selected as the final generalized level.

[0112] It should be pointed out that in the multi-granularity method, the optimistic strategy is a more relaxed condition, which must provide all possible situations based on historical data. We can select the maximum or minimum value of all possibilities according to actual needs; the pessimistic strategy is a more stringent strategy, which matches historical data and the mined rules are more accurate.

[0113] It should be noted that, in the multi-granularity method, the t value selected in the time series T={1, 2, . . . , t}, t>0 is the maximum value of the minimum rules obtained by the pessimistic strategy.

[0114] The above implementation is used to illustrate the technical solution of the present invention rather than to limit it. Those skilled in the art should understand that any modification, equivalent substitution or improvement of the technical solution of the present invention within the spirit and principle of the present invention is included in the protection scope of the present invention.

Claims

1. A method for short-term power forecasting based on generalized hierarchical multi-granularity, characterized in that: The method specifically comprises the following steps: Step S1: preprocessing and discretizing the power load data; Step S2: According to the importance of the attribute affecting the power load and the distance of the combination particle affecting the power load, select the attribute affecting the power load whose importance is greater than 0.2, wherein the attributes affecting the power load are specifically selected as follows: Step S21: Define power load The attributes that affect the power load Dependencies: Power load The attributes that affect the power load The dependency of is defined as follows: ； in, Indicates the attributes that affect the amount of power load Relative decision attribute power load The number of positive domains of Indicates the number of selected power data; Step S22: Calculate the importance of attributes that affect the power load: define the importance of attributes that affect the power load; ； in, Representation attributes In the conditional attribute Relative to the decision attribute power load The importance of Step S23: Delete the attributes that affect the power load and whose attribute importance is less than or equal to 0.2: Sort by importance and delete the attributes that affect the power load if the attribute importance is less than or equal to 0.2; Step S24: Calculate the distance between the attribute particles that affect the power load: Calculate the distance between the particles that affect the power load: in, Represents the distance between different particles. Indicates the attributes that affect the power load Exported particle structure ; Step S25: Calculate the similarity between attributes that affect the power load: Based on the different particle distances, calculate the similarity between two attributes, the formula is: in, Represents particle structure and The similarity of Represents particle structure and distance; Step S26: Filter the attribute with the smallest value of information entropy affecting power load: Select the attribute with the smallest information entropy value; Step S3: Select the attributes that affect the power load after screening, and establish a hierarchical model of the power load through a concept hierarchy tree; Step S4: According to the data of user demand and power load, first select the level with the largest attribute importance value of power load, then sort the level attributes according to their importance, select the attribute level that affects the power load, and construct the time series condition granule according to the time series, wherein the level selection specifically includes: Step S41: The first The importance of the attributes of the layer that affect the power load is as follows: in Indicates The importance of the attributes of the layer that affect the power load, Represents the hierarchical attributes that affect the power load Take the first The attribute value of the layer, other attributes take the most fine-grained attributes, Indicates the attributes that will affect the amount of power load Remove the remaining attributes that affect the power load; Step S42: The importance of attributes affecting the power load in the hierarchical model of the power load is as follows: in Represents the hierarchical attributes that affect the power load Importance, Indicates the attributes that affect the power load The weight of each level, and the sum of the weights of each level , here each level is given the same weight , , and then sort; Step S43: Construct the attribute combination granules that affect the power load on the time series day: generalize each attribute level to the coarsest granularity level with unchanged importance in descending order of the attribute importance that affects the power load, and construct the time series The conditional particle is expressed as: The timing Before Moment sequence; Step S5: using a multi-granularity method and applying different strategies to predict the power generation on the day to be predicted according to the characteristics of the predicted power load, wherein the multi-granularity method specifically includes: A multi-granularity method is used to predict the power load of the day to be predicted based on the combination of different attributes that affect the power load and matching rules; in , Respectively represent the multi-granularity method of taking optimistic and pessimistic strategies, Represents a combination of attributes that affect the power load The derived equivalence class, Represents the decision attribute power load Layer An equivalence class derived from a value.

2. The method for short-term power forecasting based on generalized hierarchical multi-granularity according to claim 1 is characterized in that: The step S1 of data preprocessing, discretization of data specifically includes: The attribute values ​​of climate, holidays, electricity price policies, emergencies and power load that affect power load are divided into discrete values ​​according to certain standards. It means that the fog levels are divided into light fog, fog, heavy fog, dense fog and very dense fog; the date types are divided into weekdays and weekends; the power load is divided into extremely low, low, normal, high and extremely high. The load value at a certain moment or the daily and monthly load is used as the decision attribute according to the demand; other attributes are discretized accordingly.

3. The method for short-term power forecasting based on generalized hierarchical multi-granularity according to claim 1 is characterized in that: The step S3 of establishing the hierarchical model specifically includes: Select attributes whose importance to the power load is greater than 0.2; Establish a conceptual hierarchical tree of power load: hierarchically cluster each attribute that affects power load and each attribute that determines power load, and establish a conceptual hierarchical tree.

4. The method for short-term power forecasting based on generalized hierarchical multi-granularity according to claim 1 is characterized in that: The multiple granularities are in a parallel relationship, and there is no inclusion relationship between the granularities.

5. The method for short-term power forecasting based on generalized hierarchical multi-granularity according to claim 1 is characterized in that: The decision attribute level selection selects the power load according to different user needs The coarse-grained level is suitable for user understanding, and the fine-grained level is more suitable for detailed decision-making.

6. The method for short-term power forecasting based on generalized hierarchical multi-granularity according to claim 1 is characterized in that: The importance of the attribute affecting the power load does not change. For the attribute affecting the power load, the level of the attribute affecting the power load is continuously generalized to calculate the importance until the level of importance changes. , then select The layer serves as the final generalization level.

7. The method for short-term power forecasting based on generalized hierarchical multi-granularity according to claim 1 is characterized in that: In the multi-granularity method, the optimistic strategy is a more relaxed condition, which must provide all possible situations based on historical data. We can select the maximum or minimum value of all possibilities according to actual needs; the pessimistic strategy is a more stringent strategy, which matches historical data and makes the mining rules more accurate; the multi-granularity method selects time series of The value is the maximum value of the minimum number of rules obtained by the pessimistic strategy.

Images