Method and system for mining wind-light-storage scheduling rule based on representation learning

Abstract

The invention discloses a wind and light storage scheduling rule mining method and system based on representation learning, and the method comprises the following steps: constructing a knowledge graph based on wind and light storage scheduling historical data; recursively decomposing a longer rule path in the knowledge graph into single rule heads by using a relation path coding model in combination with deductive performance of a logic rule, and outputting a rule score; measuring the difference between the deduced rule head and the actual target relationship by using a path closed prediction model; and evaluating and selecting high-quality rules, and constructing a scheduling rule set. The method has the characteristics of high rule mining efficiency, high interpretability and the like, and can provide powerful support for intelligent scheduling decision-making of the wind and light storage system.

Description

Technical Field

[0001] The present invention relates to a method and system for mining wind, solar and energy storage dispatching rules based on representation learning, belonging to the technical field of electric power dispatching, and used for intelligent dispatching of wind, solar and energy storage systems. Background Art

[0002] With the acceleration of global energy transformation, the proportion of renewable energy such as wind and solar energy in the power system continues to rise. However, wind and solar energy are highly intermittent and volatile, and their uncertainty poses a huge challenge to the stability of the power grid. In order to make full use of wind and solar resources and ensure the safe operation of the power grid, wind, solar and storage complementary technology has gradually become the focus of power dispatching research. By jointly dispatching wind, solar and energy storage systems, this technology can effectively alleviate the impact of renewable energy generation fluctuations on the power grid and enhance the flexibility and reliability of the system. However, the wind, solar and storage combined power generation system still faces many challenges, such as uncontrollable output power, time-varying electricity prices, and random fluctuations in new energy output. Traditional dispatching methods usually rely on fixed dispatching plans, which are difficult to flexibly respond to the randomness of new energy output, resulting in low dispatching efficiency and failure to fully tap the potential of new energy.

[0003] Knowledge graph is a way of expressing knowledge in the form of graph structure to represent entities and their relationships, which can help machines better understand and process complex data. Knowledge graph rule mining is the process of automatically discovering potential logical rules from graphs, usually by analyzing the paths between entities, logical reasoning or representation learning to generate high-quality rules. In the field of wind, solar and energy storage scheduling, knowledge graphs can represent the entities and their relationships in wind, solar and energy storage systems in the form of graph structures, helping intelligent systems understand and process complex scheduling data. By mining rules on wind, solar and energy storage scheduling knowledge graphs, potential scheduling rules in wind, solar and energy storage systems can be automatically discovered. These potential scheduling rules can effectively integrate the characteristics of wind, solar and energy storage systems, and dynamically respond to the volatility and uncertainty of energy output. This helps to improve the efficiency of scheduling decisions of wind, solar and energy storage dispatchers and reduce reliance on traditional experience. At the same time, the mined rules help to achieve intelligent scheduling optimization, enabling the system to automatically adjust the scheduling plan under changing energy demand and power generation conditions, ensuring the maximum utilization of energy resources and the stable operation of the power grid.

[0004] However, existing rule mining methods based on knowledge graphs ignore the deductive nature of logical rules and directly search for rule instances in knowledge graphs, which results in a huge search space and reduces the efficiency of rule mining. In addition, these methods are difficult to mine rules that lack instance support, which limits their application in complex scenarios. Summary of the invention

[0005] Purpose of the invention: In view of the problems and shortcomings existing in the prior art, the present invention provides a method and system for mining wind, solar and storage scheduling rules based on representation learning to improve the efficiency and interpretability of rule mining, thereby providing strong support for intelligent scheduling decisions of wind, solar and storage systems.

[0006] Technical solution: A method for mining wind, solar and energy storage dispatching rules based on representation learning, including the following steps:

[0007] (1) Data preprocessing: Based on the historical data of wind, solar and energy storage scheduling, we extract scheduling-related entities and relationships and build a knowledge graph;

[0008] (2) Relational path encoding: Combining the deductive nature of logical rules, the relational path encoding model is used to recursively decompose the longer rule paths in the knowledge graph into single rule heads and output the rule scores;

[0009] (3) Path closure prediction: The path closure prediction model is used to measure the gap between the derived rule head and the actual target relationship;

[0010] (4) Candidate rule evaluation: Based on the semantic consistency between the rule body and the rule header, the candidate rules are evaluated and high-quality rules are selected to build a scheduling rule set.

[0011] Furthermore, the specific steps of relationship path encoding in step (2) are as follows:

[0012] (2.1) Use the random walk algorithm to sample paths from the knowledge graph constructed in step (1).

[0013] (2.2) Given a fixed window size s = 2, by The sliding window with a size of s is used to generate a relationship pair. When processing a local sliding window (i.e., a relationship pair), the relationship within the sliding window strictly follows the chain structure. The sequence model RNN is used to encode the sliding window and convert the sliding window relationship pair w i =[r i ,r i+1 ] as input, RNN outputs:

[0014]

[0015] in, is the predicate r i In w i The hidden state in represents a d-dimensional real number space, is the predicate r i+1 In w i The hidden state in .

[0016] (2.3) Final hidden state It is usually used to represent the entire sequence, so the relationship between w i It can be expressed as Used to determine the relationship between [r i ,r i+1 ] are likely to appear at the same time. If this set of relations always appears in pairs, it means that they have a higher probability of forming a meaningful combination. In order to determine the specific order of performing relation replacement operations, that is, to select specific relation pairs for operation in each step of derivation, the best replacement relation pair can be selected according to the following formula:

[0017]

[0018] Among them, f(·) represents a fully connected neural network, which can be represented from the vector Learn and determine the effective probability of the relationship pair in the sliding window. μ is a probability distribution calculated by the softmax function, indicating the possibility of each sliding window forming a meaningful combination. The probability μ i The highest w i As the best replacement relationship pair.

[0019] (2.4) When deriving the relationship at each step, select μ i The highest w i Use the Multi-layer Perceptron (MLP) classifier to calculate the relationship between each relationship r in the knowledge graph relationship set. k ∈R can replace the probability q(r k |r i ,r j ). MLP classifier is a two-layer fully connected neural network with parameter θ, using softmax as the activation function of the last layer. The MLP classifier classifies the relationship r i and r j The vector representation of is taken as input, and the final output relation pair is represented by each relation r in the knowledge graph relation set. k ∈R replacement probability, that is, q(r k |r i ,r j ), the formula is as follows:

[0020]

[0021] Among them, r k is a relation in the knowledge graph relation set, q(r k |r i ,r j ) represents the relationship r kYou can replace the relation pair [r i ,r j ] probability.

[0022] (2.5) In each recursive replacement of the relationship path, select the relationship with the largest q(r h |r i ,r j ) of a single relation r h Replace the relationship path. By introducing All relationship vectors are represented as weighted aggregations to reduce computational complexity. Given a relationship path It can be calculated as follows:

[0023]

[0024]

[0025] in, is a single relation r k Replacement relationship pair The weighted average representation of .

[0026] Furthermore, the specific steps of closed path prediction in step (3) are as follows: using the path closure prediction model p(r t |r h ) measures the ideal prediction r in the process of logical rule induction t and real observation r h The gap between them is as follows:

[0027]

[0028] Among them, p(r t |r b ) represents the target relationship r t The relationship path r b The path closure prediction model p(r t |r h ) is a two-layer fully connected neural network, the inter-layer activation function is ReLU, and the last layer activation function is sigmoid.

[0029] Furthermore, the specific steps of candidate rule evaluation in step (4) are as follows:

[0030] (4.1) The rule scoring function is used to calculate the probability that the rule body can be replaced by the rule head based on semantic consistency, and the quality of the candidate rule is evaluated to replace the confidence that requires a large number of rule instances to support it. The formula is as follows:

[0031] s(r h , r b)=q(r h =r i |r b ),r i ∈R

[0032] Among them, s(r h ,r b ) indicates the rule header r h Replace rule body r b The possibility of q(r h =r i |r b ) is a probability function based on semantic consistency judgment learned by RNN.

[0033] (4.2) Sort all candidate rules by rule scoring and select the top k rules with the highest scores as high-quality rules.

[0034] (4.3) The selected high-quality rules are stored in a relational database, using a nested structure to represent and store them. The storage structure of each rule can be expressed as:

[0035] 〈RuleID, Body, Head, Attributes>

[0036] Among them, RuleID represents the unique identifier of the rule, Body is the rule body stored in the form of an ordered list, Head represents the rule header, and Attributes represents additional attributes such as rule score and timestamp.

[0037] A wind, solar and energy storage dispatching rule mining system based on representation learning, including the following modules:

[0038] (1) Data preprocessing module: Based on the historical data of wind, solar and energy storage scheduling, it extracts scheduling-related entities and relationships and constructs a knowledge graph;

[0039] (2) Relationship path encoding module: Combined with the deductive nature of logical rules, the relationship path encoding model is used to recursively decompose the longer rule paths in the knowledge graph into single rule heads and output the rule scores;

[0040] (3) Closed path prediction module: uses the path closure prediction model to measure the gap between the derived rule head and the actual target relationship;

[0041] (4) Candidate rule evaluation module: Evaluate candidate rules based on the semantic consistency between the rule body and the rule header, select high-quality rules, and build a scheduling rule set.

[0042] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, a method for realizing a wind, solar, and energy storage scheduling rule mining system based on representation learning as described above is implemented.

[0043] A computer-readable storage medium stores a computer program for executing a method for implementing a wind-solar-storage scheduling rule mining system based on representation learning as described above.

[0044] Beneficial effects: The wind, solar, and storage scheduling rule mining method and system based on representation learning provided by the present invention significantly reduces the dependence on rule instances, and can directly mine high-quality scheduling rules when rule instances are insufficient through pattern learning. The present invention combines the deductive nature of logical rules and recursively decomposes relationship paths to effectively narrow the search space and improve the efficiency and accuracy of rule mining. At the same time, the rules are evaluated based on semantic consistency to improve the reliability of the rules. The high-quality rules mined from the historical data of wind, solar, and storage scheduling can provide strong support for the intelligent scheduling decisions of wind, solar, and storage systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a model training framework diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The present invention is further explained below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0047] A method for implementing a wind, solar and energy storage scheduling rule mining system based on representation learning specifically comprises the following steps:

[0048] Step (1) Data preprocessing: Based on the historical data of wind, solar and energy storage scheduling, extract scheduling-related entities and relationships and construct a knowledge graph.

[0049] Taking the historical data of wind, solar and energy storage dispatch in a certain province as an example, the results of extracting dispatch-related entities and relations are shown in Table 1.

[0050] Table 1. Partial extraction results of scheduling-related entities and relations

[0051]

[0052] Relation path encoding in step (2): Combined with the deductive nature of logical rules, the relation path encoding model is used to recursively decompose the longer rule paths in the knowledge graph into single rule heads and output the rule scores;

[0053] (2.1) Use the random walk algorithm to sample paths from the knowledge graph constructed in step (1).

[0054] (2.2) Given a fixed window size s = 2, by The sliding window with a size of s is used to generate a relationship pair. When processing a local sliding window (i.e., a relationship pair), the relationship within the sliding window strictly follows the chain structure. The sequence model RNN is used to encode the sliding window and convert the sliding window relationship pair w i =[r i ,r i+1 ] as input, RNN outputs:

[0055]

[0056] in, is the predicate r i In w i The hidden state in is the predicate r i+1 In w i The hidden state in .

[0057] (2.3) Final hidden state It is usually used to represent the entire sequence, so the relationship between w i It can be expressed as Used to determine the relationship between [r i ,r i+1 ] are likely to appear at the same time. If this set of relations always appears in pairs, it means that they have a higher probability of forming a meaningful combination. In order to determine the specific order of performing relation replacement operations, that is, to select specific relation pairs for operation in each step of derivation, the best replacement relation pair can be selected according to the following formula:

[0058]

[0059] Among them, f(·) represents a fully connected neural network, which can be represented from the vector Learn and determine the effective probability of the relationship pair in the sliding window. μ is a probability distribution calculated by the softmax function, which indicates the possibility of each sliding window forming a meaningful combination. i The highest w i As the best replacement relationship pair.

[0060] (2.4) When deriving the relationship at each step, select μ i The highest w iUse the Multi-layer Perceptron (MLP) classifier to calculate the relationship between each relationship r in the knowledge graph relationship set. k ∈R can replace the probability q(r k |r i ,r j ). MLP classifier is a two-layer fully connected neural network with parameter θ, using softmax as the activation function of the last layer. The MLP classifier classifies the relationship r i and r j The vector representation of is taken as input, and the final output relation pair is represented by each relation r in the knowledge graph relation set. k ∈R replacement probability, that is, q(r k |r i ,r j ), the formula is as follows:

[0061]

[0062] Among them, r k is a relation in the knowledge graph relation set, q(r k |r i ,r j ) represents the relationship r k You can replace the relation pair [r i ,r j ] probability.

[0063] (2.5) In each recursive replacement of the relationship path, select the relationship with the largest q(r h |r i ,r j ) of a single relation r h Replace the relationship path. By introducing All relationship vectors are represented as weighted aggregations to reduce computational complexity. Given a relationship path It can be calculated as follows:

[0064]

[0065]

[0066] in, is a single relation r k Replacement relationship pair The weighted average representation of .

[0067] Step (3) closed path prediction: Through the path closure prediction model p(r t |r h) measures the ideal prediction r in the process of logical rule induction t and real observation r h The gap between them is as follows:

[0068]

[0069] Among them, p(r t |r b ) represents the target relationship r t The relationship path r b The path closure prediction model p(r t |r h ) is a two-layer fully connected neural network, the inter-layer activation function is ReLU, and the last layer activation function is sigmoid.

[0070] Step (4) Candidate rule evaluation: Based on the semantic consistency between the rule body and the rule header, candidate rules are evaluated and high-quality rules are selected to construct a scheduling rule set.

[0071] (4.1) The rule scoring function is used to calculate the probability that the rule body can be replaced by the rule head based on semantic consistency, and the quality of the candidate rule is evaluated to replace the confidence that requires a large number of rule instances to support it. The formula is as follows:

[0072] s(r h , r b )=q(r h =r i |r b ), r i ∈R

[0073] Among them, s(r h ,r b ) indicates the rule header r h Replace rule body r b The possibility of q(r h =r i |r b ) is a probability function based on semantic consistency judgment learned by RNN.

[0074] (4.2) Sort all candidate rules by rule scoring and select the top k rules with the highest scores as high-quality rules.

[0075] Some of the high-quality scheduling rules mined are shown in Table 2.

[0076] Table 2 Some high-quality scheduling rules mined

[0077]

[0078] (4.3) The selected high-quality rules are stored in a relational database, using a nested structure to represent and store them. The storage structure of each rule can be expressed as:

[0079] 〈RuleID, Body, Head, Attributes〉

[0080] Among them, RuleID represents the unique identifier of the rule, Body is the rule body stored in the form of an ordered list, Head represents the rule header, and Attributes represents additional attributes such as rule score and timestamp.

[0081] The wind, solar and energy storage dispatching rule mining system based on representation learning includes the following modules:

[0082] (1) Data preprocessing module: Based on the historical data of wind, solar and energy storage scheduling, it extracts scheduling-related entities and relationships and constructs a knowledge graph;

[0083] (2) Relationship path encoding module: Combined with the deductive nature of logical rules, the relationship path encoding model is used to recursively decompose the longer rule paths in the knowledge graph into single rule heads and output the rule scores;

[0084] (3) Closed path prediction module: uses the path closure prediction model to measure the gap between the derived rule head and the actual target relationship;

[0085] (4) Candidate rule evaluation module: Evaluate candidate rules based on the semantic consistency between the rule body and the rule header, select high-quality rules, and build a scheduling rule set.

[0086] The implementation process and method of the system are the same.

[0087] Obviously, those skilled in the art should understand that the various steps of the wind-solar-storage scheduling rule mining method based on representation learning or the various modules of the wind-solar-storage scheduling rule mining system based on representation learning of the above-mentioned embodiments of the present invention can be implemented by a general-purpose computing device, which can be concentrated on a single computing device or distributed on a network composed of multiple computing devices. Optionally, they can be implemented with program codes executable by the computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order than here, or they can be made into individual integrated circuit modules respectively, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. In this way, the embodiments of the present invention are not limited to any specific combination of hardware and software.

Claims

1. A method for mining wind, solar and energy storage scheduling rules based on representation learning, characterized in that: The following steps are involved: (1) Data preprocessing: Based on the historical data of wind, solar and energy storage scheduling, we extract scheduling-related entities and relationships and build a knowledge graph; (2) Relational path encoding: Combining the deductive nature of logical rules, the relational path encoding model is used to recursively decompose the rule path in the knowledge graph into a single rule head and output the rule score; (3) Path closure prediction: The path closure prediction model is used to measure the gap between the derived rule head and the actual target relationship; (4) Candidate rule evaluation: Based on the semantic consistency between the rule body and the rule header, the candidate rules are evaluated and high-quality rules are selected to build a scheduling rule set.

2. The method for mining wind, solar and energy storage scheduling rules based on representation learning according to claim 1, characterized in that: The specific steps of relationship path encoding in step (2) are as follows: (2.1) Use the random walk algorithm to sample paths from the knowledge graph; (2.2) Given a window size s, generate a relation pair by sliding a window of size s on the relation path; When processing a local sliding window, the relationship within the sliding window follows a chain structure; the sequence model RNN is used to encode the sliding window and the sliding window relationship pair w i =[r i ,r i+1 ] as input, RNN outputs: in, is the predicate r i In w i The hidden state in is the predicate r i+1 In w i The hidden state in ; (2.3) Final hidden state Usually used to represent the entire sequence, the relationship between w i Expressed as Used to determine the relationship between [r i ,r i+1 ]The possibility of simultaneous occurrence; the best replacement relationship pair is selected according to the following formula: Among them, f(·) represents a fully connected neural network, which can be represented from the vector Learn and determine the effective probability of the relationship pair in the sliding window; μ is a probability distribution calculated by the softmax function, indicating the possibility of each sliding window forming a meaningful combination, the probability μ i The highest w i As the best replacement relationship pair; (2.4) When deriving the relationship at each step, select μ i The highest w i As the best replacement relation pair, the relation is replaced; the multi-layer perceptron classifier is used to calculate each relation r in the knowledge graph relation set. k ∈R can replace the probability q(r k |r i ,r j ); MLP classifier It is a two-layer fully connected neural network with parameter θ, using softmax as the activation function of the last layer; the MLP classifier classifies the relationship r i and r j The vector representation of is taken as input, and the final output relation pair is represented by each relation r in the knowledge graph relation set. k ∈R replacement probability, that is, q(r k |r i ,r j ), the formula is as follows: Among them, r k is a relation in the knowledge graph relation set, q(r k |r i ,r j ) represents the relationship r k You can replace the relation pair [r i ,r j ]; (2.5) In each recursive replacement of the relationship path, select the relationship with the largest q(r h |r i ,r j ) of a single relation r h Replace the relationship path; by introducing All relationship vectors represent weighted aggregation; given a relationship path It is calculated as follows: in, is a single relation r k Replacement relationship pair The weighted average representation of .

3. The method for mining wind, solar and energy storage scheduling rules based on representation learning according to claim 1 is characterized in that: The specific steps of closed path prediction in step (3) are as follows: t |r h ) measures the ideal prediction r in the process of logical rule induction t and real observation r h The gap between them is as follows: Among them, p(r t |r b ) represents the target relationship r t The relationship path r b Closure ratio; path closure prediction model p(r t |r h ) is a two-layer fully connected neural network, the inter-layer activation function is ReLU, and the last layer activation function is sigmoid.

4. The method for mining wind, solar and energy storage scheduling rules based on representation learning according to claim 1 is characterized in that: The specific steps of candidate rule evaluation in step (4) are as follows: (4.1) The rule scoring function is used to calculate the probability that the rule body can be replaced by the rule head based on semantic consistency, and the quality of the candidate rule is evaluated to replace the confidence that requires a large number of rule instances to support it. The formula is as follows: s(r h ,r b )=q(r h =r i |r b ),r i ∈R Among them, s(r h ,r b ) indicates the rule header r h Replace rule body r b The possibility of q(r h =r i |r b ) is a probability function based on semantic consistency judgment learned by RNN; (4.2) Sort all candidate rules by rule scoring and select the top k rules with the highest scores as high-quality rules; (4.3) The selected high-quality rules are stored in a relational database, using a nested structure to represent and store them; the storage structure of each rule is represented as follows: <RuleID,Body,Head,Attributes> Among them, RuleID represents the unique identifier of the rule, Body is the rule body stored in the form of an ordered list, Head represents the rule header, and Attributes represents additional attributes.

5. A wind, solar and energy storage scheduling rule mining system based on representation learning, characterized in that: Includes the following modules: (1) Data preprocessing module: Based on the historical data of wind, solar and energy storage scheduling, it extracts scheduling-related entities and relationships and constructs a knowledge graph; (2) Relationship path encoding module: Combined with the deductive nature of logical rules, the relationship path encoding model is used to recursively decompose the longer rule paths in the knowledge graph into single rule heads and output the rule scores; (3) Closed path prediction module: uses the path closure prediction model to measure the gap between the derived rule head and the actual target relationship; (4) Candidate rule evaluation module: Evaluate candidate rules based on the semantic consistency between the rule body and the rule header, select high-quality rules, and build a scheduling rule set.

6. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for implementing a wind, solar, and storage scheduling rule mining system based on representation learning as described in any one of claims 1 to 4 is implemented.

7. A computer-readable storage medium storing a computer program for executing a method for implementing a wind-solar-storage scheduling rule mining system based on representation learning as described in any one of claims 1 to 4.

Images