Method for improving analysis effect of large model in photovoltaic service field based on FHI logic derivation framework

Abstract

The invention provides a method for improving an analysis effect of a large model in a photovoltaic service field based on an FHI logic derivation framework. The method comprises the following steps: establishing an FHI logic reasoning basic model; for a target problem in the photovoltaic business field, repeatedly calling the FHI logical reasoning basic model to obtain an FHI logical reasoning model, and then converting the FHI logical reasoning model into a argument fact tree; and analyzing and processing each Fact fact sub-node in the argument fact tree to obtain an embedding embedded code embedded with hierarchical features in the argument fact tree, and inputting the embedding embedded code into a transformer module. According to the method, a large language model in the field of photovoltaic business not only understands photovoltaic professional knowledge, but also can realize long-link reasoning analysis based on a reasoning logic chain and a scheduling information system business interface, so that the reasoning process of the large language model conforms to business logic, the understanding depth of the large language model on the photovoltaic business is enhanced, and the understanding efficiency of the large language model on the photovoltaic business is improved. Therefore, the accuracy and reliability of the prediction analysis result of the large language model are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of big data analysis and processing, and specifically relates to a method for improving the analysis effect of a large model in the photovoltaic business field based on a FHI logic deduction framework. Background Art

[0002] In the photovoltaic business field, various large language models are often needed according to application requirements, such as: large language models for evaluating the health status of photovoltaic equipment; large language models for predicting the power and power generation of photovoltaic equipment by combining meteorological and topographic data.

[0003] In the prior art, the large language model in the photovoltaic business field is a general pre-trained large language model: by pre-training on massive data, the common knowledge of language is learned to improve the transfer learning ability and generalization ability on general tasks.

[0004] However, due to the complex equipment link relationships in the photovoltaic business field, the use of general pre-trained large language models has weak processing capabilities for multi-hop tasks in long logical chains, resulting in low accuracy of prediction results. Summary of the invention

[0005] In view of the defects of the prior art, the present invention provides a method for improving the analysis effect of large models in the photovoltaic business field based on the FHI logic deduction framework, which can effectively solve the above problems.

[0006] The technical solution adopted by the present invention is as follows:

[0007] The present invention provides a method for improving the analysis effect of a large model in the photovoltaic business field based on the FHI logic deduction framework, comprising the following steps:

[0008] Step S1, establishing a basic model of FHI logic reasoning; the basic model of FHI logic reasoning is a tree structure, including a Fact node, a Hypothesize node, an Investigate node and a Fact sub-node connected hierarchically from top to bottom;

[0009] Step S2, for the target problem in the photovoltaic business field, repeatedly call the FHI logical reasoning basic model to obtain the FHI logical reasoning model; the method is:

[0010] After the FHI logic reasoning basic model matching the target problem in the photovoltaic business field is established, the associated business knowledge and business logic are called, and each of the Fact fact sub-nodes is analyzed to determine whether the Fact fact sub-node is the bottom-level node associated with the target problem in the photovoltaic business field in the business logic. If so, the re-calling of the FHI logic reasoning basic model is terminated; if not, the Fact fact sub-node is used as the Fact fact node, and the FHI logic reasoning basic model is called again to increase the level of the tree structure. If the cycle is continuous, the FHI logic reasoning model is finally established;

[0011] Step S3, converting the FHI logical reasoning model into an argument fact tree; the argument fact tree only has Fact fact nodes and Fact fact sub-nodes of each level;

[0012] Step S4, analyzing and processing each of the Fact sub-nodes in the argument fact tree to obtain an embedding code embedded with the hierarchical features in the argument fact tree;

[0013] Step S5, embedding and encoding each of the embeddings having a hierarchical cascade relationship, sorting them by level, and inputting them into a transformer module, and the transformer module outputs an analysis result for the target problem in the photovoltaic business field.

[0014] Preferably, in step S1, the process of establishing the FHI logical reasoning basic model is:

[0015] Step S1.1, storing the semantic description information of the target problem into the Fact node;

[0016] Step S1.2, calling the business knowledge and business logic associated with the target problem, analyzing and reasoning the semantic description information of the target problem stored in the Fact node, obtaining N rational hypothesis texts associated with the target problem, and for each rational hypothesis text, generating a Hypothesize rational hypothesis node as a child node of the Fact node;

[0017] Step S1.3, for each of the Hypothesize rational hypothesis nodes, obtain M associated Investigate system investigation nodes and use them as child nodes of the Hypothesize rational hypothesis node; each of the Investigate system investigation nodes is bound to a model agent, and by executing the model agent, text-based argument facts are obtained;

[0018] Step S1.4, the text-type argument facts of each Investigate system investigation node are stored in the Fact fact child node which is a child node of the Investigate system investigation node;

[0019] From top to bottom, the FHI logical reasoning basic model is established to obtain a tree structure formed by Fact nodes, Hypothesize nodes, Investigate nodes and Fact sub-nodes.

[0020] Preferably, the business knowledge and business logic are a large language model enhanced based on knowledge graph and vector retrieval.

[0021] Preferably, the model agent bound to the Investigate system investigation node is a model agent obtained by training with historical data;

[0022] The training method is:

[0023] The model agent is trained using key-value pairs in the form of {"text":(model agent name, key parameter list)}, so that the model agent has the capabilities of semantic understanding, top_N model agent recall, and key parameter extraction.

[0024] Preferably, the model agent has an input processing submodule, a business interface calling submodule and an integrated output submodule;

[0025] The input processing submodule is used to perform semantic understanding on the rational hypothesis text input from the Hypothesize rational hypothesis node;

[0026] The business interface calling submodule is used to call the associated business database or business data table according to the semantic understanding of the input processing submodule, and query and obtain key parameters associated with the semantic understanding as query results;

[0027] The integration output submodule is used to form a key parameter table with the multiple key parameters obtained by the business interface calling submodule, and output it to the Fact fact subnode.

[0028] Preferably, in step S3, the FHI logical reasoning model is converted into an argument fact tree, specifically:

[0029] Delete each Hypothesize rational hypothesis node and Investigate system investigation node in the FHI logical reasoning model, keep the number of levels and cascade relationship of each Fact fact sub-node unchanged, and obtain the argument fact tree; wherein the number of levels and cascade relationship are called argument derivation level relationship.

[0030] Preferably, step S4 is specifically:

[0031] The Fact sub-node stores text-based argument facts and argument derivation hierarchical relationships;

[0032] Segmenting the text-like argument facts, wherein the text-like argument facts have M-dimensional segmentations, and each segmentation is converted into a digital representation of semantics, thereby obtaining an M-dimensional text semantic vector v_semantic;

[0033] According to the text position information of each word in the text class argument fact, an M-dimensional text position vector v_position is obtained by combining sine and cosine with different periods;

[0034] Relying on the argument fact tree, the hierarchical relationship is deduced according to the arguments of the Fact fact sub-node, including the hierarchical level and the order within the level, to form an initial hierarchical logic vector with a dimension of D, and then a structural position vector v_structure with a dimension of M is obtained through linear layer mapping;

[0035] The text semantic vector v_semantic, the text position vector v_position and the structure position vector v_structure are added together to obtain the embedding encoding embeding of the text class argument fact.

[0036] Preferably, step S5 is specifically:

[0037] The transformer module includes an encoder, a decoder, a linear layer and a Softmax layer;

[0038] Embedding the embedding codes in a hierarchical cascade relationship, sorting them by level to form an embedding code sequence, and inputting the embedding code sequence into the encoder;

[0039] The encoder captures the global dependencies between the elements in the embedding coding sequence input, generates a series of context-aware representations, and inputs them to the decoder;

[0040] The decoder is used to generate an output sequence based on the context-aware representation;

[0041] The linear layer is used to map the output sequence of the decoder to a vector;

[0042] The Softmax layer is used to convert the vector output by the linear layer into a probability distribution.

[0043] The method provided by the present invention for improving the analysis effect of a large model in the photovoltaic business field based on the FHI logic deduction framework has the following advantages:

[0044] The present invention provides a method for improving the analysis effect of a large model in the photovoltaic business field based on the FHI logic deduction framework, so that the large language model in the photovoltaic business field not only understands photovoltaic professional knowledge, but also can realize long-link reasoning analysis based on the reasoning logic chain and the business interface of the scheduling information system, so that the reasoning process of the large language model conforms to the business logic, and enhances the depth of understanding of the photovoltaic business by the large language model, thereby improving the accuracy and reliability of the prediction and analysis results of the large language model. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A schematic diagram of the FHI logical reasoning model provided by the present invention for generating an argument fact tree;

[0046] Figure 2 The schematic diagram of the transformer module enhanced based on the argument fact tree provided by the present invention;

[0047] Figure 3 An example diagram of generating an FHI logical reasoning model in the photovoltaic business field provided by an embodiment of the present invention;

[0048] Figure 4 An example diagram of an argument fact tree generated according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to make the technical problems, technical solutions and beneficial effects solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0050] The present invention provides a method for improving the analysis effect of a large model in the photovoltaic business field based on the FHI logic deduction framework, so that the large language model in the photovoltaic business field not only understands photovoltaic professional knowledge, but also can realize long-link reasoning analysis based on the reasoning logic chain and the business interface of the scheduling information system, so that the reasoning process of the large language model conforms to the business logic, and enhances the depth of understanding of the photovoltaic business by the large language model, thereby improving the accuracy and reliability of the prediction and analysis results of the large language model.

[0051] The overall process of the present invention can be divided into the following two parts:

[0052] Based on the FHI logical reasoning model of Fact-Rational Hypothesis-Systematic Investigation (FHI), the argument fact tree is obtained through the cyclic scheduling of semantic understanding and model agent;

[0053] By utilizing the hierarchical structure information of the argument fact tree, the logical relationship in the position encoding is enhanced, the transformer-type model's understanding of specific photovoltaic business issues is improved, and reasoning and analysis are achieved.

[0054] The present invention is described in detail below:

[0055] The present invention provides a method for improving the analysis effect of a large model in the photovoltaic business field based on the FHI logic deduction framework, comprising the following steps:

[0056] Step S1, establishing a basic model of FHI logical reasoning;

[0057] Combination Figure 1 The FHI logical reasoning basic model is a tree structure, including Fact nodes, Hypothesize nodes, Investigate nodes and Fact sub-nodes connected hierarchically from top to bottom;

[0058] The establishment process of the FHI logical reasoning basic model is as follows:

[0059] Step S1.1, storing the semantic description information of the target problem into the Fact node;

[0060] Step S1.2, calling the business knowledge and business logic associated with the target problem, analyzing and reasoning the semantic description information of the target problem stored in the Fact node, obtaining N rational hypothesis texts associated with the target problem, and for each rational hypothesis text, generating a Hypothesize rational hypothesis node as a child node of the Fact node;

[0061] In this step, the business knowledge and business logic are a large language model enhanced based on knowledge graph and vector retrieval.

[0062] Step S1.3, for each of the Hypothesize rational hypothesis nodes, obtain M associated Investigate system investigation nodes and use them as child nodes of the Hypothesize rational hypothesis node; each of the Investigate system investigation nodes is bound to a model agent, and by executing the model agent, text-based argument facts are obtained;

[0063] In this step, the model agent bound to the Investigate system investigation node is a model agent obtained by training with historical data;

[0064] The training method is:

[0065] The model agent is trained using key-value pairs in the form of {"text":(model agent name, key parameter list)}, so that the model agent has the capabilities of semantic understanding, top_N model agent recall, and key parameter extraction.

[0066] The model agent comprises an input processing submodule, a business interface calling submodule and an integrated output submodule;

[0067] The input processing submodule is used to perform semantic understanding on the rational hypothesis text input from the Hypothesize rational hypothesis node;

[0068] The business interface calling submodule is used to call the associated business database or business data table according to the semantic understanding of the input processing submodule, and query and obtain key parameters associated with the semantic understanding as query results;

[0069] The integration output submodule is used to form a key parameter table with the multiple key parameters obtained by the business interface calling submodule, and output it to the Fact fact subnode.

[0070] Step S1.4, the text-type argument facts of each Investigate system investigation node are stored in the Fact fact child node which is a child node of the Investigate system investigation node;

[0071] From top to bottom, the FHI logical reasoning basic model is established to obtain a tree structure formed by Fact nodes, Hypothesize nodes, Investigate nodes and Fact sub-nodes.

[0072] Therefore, in the present invention, the FHI logical reasoning basic model is a logical reasoning architecture that starts from facts, goes through model semantic understanding and model agent scheduling, and finally returns to facts, including Fact fact nodes, Hypothesize rational hypothesis nodes, Investigate system investigation nodes and Fact fact sub-nodes connected from top to bottom.

[0073] Fact node is used to store an objective fact state, including the description of the problem and the supporting facts obtained based on reasoning investigation;

[0074] The Hypothesize rational hypothesis node is a rational hypothesis made based on the facts in the parent Fact fact node, combined with business knowledge and logic. It is the ideological guidance for the next step of verification work;

[0075] The Investigate system investigation node is a verification of the parent Hypothesize rational hypothesis node, which involves the scheduling selection of the proprietary model agent and the logical concatenation of the sub-steps within the model agent;

[0076] When the Investigate system investigation node realizes the interaction with the real world and real business information system through a proprietary model agent, the final result will be stored in its child Fact sub-node.

[0077] A Fact node can correspond to one or more Hypothesize nodes, a Hypothesize node can correspond to one or more Investigate nodes, and an Investigate node can only correspond to one Fact child node. The root node of the complete Fact-Hypothesis-Investigation (FHI) logical reasoning architecture must be a Fact node, and all terminal nodes (nodes without child nodes) must be Fact child nodes.

[0078] From the Fact node to the Hypothesize rational hypothesis node, it is an assumption based on the parent Fact node under the constraints of business knowledge and logic. Its implementation process can be achieved through a large model efficiency improvement method based on knowledge graph and vector retrieval enhancement. Specifically, this method generates a vector expert library from the professional knowledge in the photovoltaic field, and generates a knowledge graph from the spatiotemporal association of photovoltaic equipment and status, so as to improve the application effect of the general large language model in the photovoltaic power field. Through this step, "generating rational hypothesis text that conforms to business logic based on fact semantics" is achieved.

[0079] From the Hypothesize rational hypothesis node to the Investigate system investigation node, the purpose is to select one or more appropriate model agents based on the semantic understanding of the rational hypothesis text of the parent node. The implementation process is: construct a key-value pair of {"text": (model agent name, key parameter list)} from the historical scheduling record for model agent training, so that the model has the ability to understand the hypothesis semantics, recall the top_N model agents, and extract key parameters.

[0080] From the Investigate system investigation node to the Fact fact sub-node, it relies on the specific implementation and actual execution of each model agent. In the model agent construction part, each model agent contains three sub-modules: "input processing", "business interface call" and "integration output"; a complete model agent is composed of one or more sub-modules in series; the model agent construction part relies on manual coding. In the actual execution part, the key parameter list is used as the output of the model agent, and the operation is realized automatically.

[0081] Step S2, for the target problem in the photovoltaic business field, repeatedly call the FHI logical reasoning basic model to obtain the FHI logical reasoning model; the method is:

[0082] After the FHI logic reasoning basic model matching the target problem in the photovoltaic business field is established, the associated business knowledge and business logic are called, and each of the Fact fact sub-nodes is analyzed to determine whether the Fact fact sub-node is the bottom-level node associated with the target problem in the photovoltaic business field in the business logic. If so, the re-calling of the FHI logic reasoning basic model is terminated; if not, the Fact fact sub-node is used as the Fact fact node, and the FHI logic reasoning basic model is called again to increase the level of the tree structure. If the cycle is continuous, the FHI logic reasoning model is finally established;

[0083] As an implementation method, by continuously calling the FHI logical reasoning basic model, a logical reasoning tree FHI-TREE with increasing levels is obtained. When the Hypothesize rational hypothesis is exhausted or a closed loop appears in the logical reasoning tree FHI-TREE or the set maximum iteration depth is reached, the generation is stopped to obtain the FHI logical reasoning model.

[0084] Step S3, converting the FHI logical reasoning model into an argument fact tree; combining Figure 1 , the argument fact tree only has Fact fact nodes and the Fact fact sub-nodes of each level;

[0085] In this step, the FHI logical reasoning model is converted into an argument fact tree, specifically:

[0086] Delete each Hypothesize rational hypothesis node and Investigate system investigation node in the FHI logical reasoning model, keep the number of levels and cascade relationship of each Fact fact sub-node unchanged, and obtain the argument fact tree; wherein the number of levels and cascade relationship are called argument derivation level relationship.

[0087] In the specific implementation, each Hypothesize rational hypothesis node and Investigate system investigation node in the FHI logical reasoning model is removed in turn, and a new connection is built between the original "grandfather node" and the original "grandchild node". When only the Fact fact node and the Fact fact child node are retained in the tree, it is recorded as the "argument fact tree".

[0088] Step S4, analyzing and processing each of the Fact sub-nodes in the argument fact tree to obtain an embedding code embedded with the hierarchical features in the argument fact tree;

[0089] This step combines Figure 2 , specifically:

[0090] The Fact sub-node stores text-based argument facts and argument derivation hierarchical relationships;

[0091] Segmenting the text-like argument facts, wherein the text-like argument facts have M-dimensional segmentations, and each segmentation is converted into a digital representation of semantics, thereby obtaining an M-dimensional text semantic vector v_semantic;

[0092] According to the text position information of each word in the text class argument fact, an M-dimensional text position vector v_position is obtained by combining sine and cosine with different periods;

[0093] Relying on the argument fact tree, the hierarchical relationship is deduced according to the arguments of the Fact fact sub-node, including the hierarchical level and the order within the level, to form an initial hierarchical logic vector with a dimension of D, and then a structural position vector v_structure with a dimension of M is obtained through linear layer mapping;

[0094] The text semantic vector v_semantic, the text position vector v_position and the structure position vector v_structure are added together to obtain the embedding encoding embeding of the text class argument fact.

[0095] Specifically, the argument fact tree contains two parts of information: each node contains text-type argument facts, and the relationship between nodes contains the argument derivation hierarchical relationship.

[0096] The Transformer model uses an attention mechanism to capture long-distance dependencies in sequence data, abandoning the recursive and convolutional structures in traditional recurrent neural networks (RNNs) and convolutional neural networks (CNNs). Since the Transformer model does not contain any recurrent structure, it cannot capture the order information of elements in the sequence. Position encoding is added to the embedding, allowing the Transformer model to understand the position of elements in the sequence.

[0097] Traditionally, the process of converting text segmentation embedding into Embedding is as follows: first, the text is segmented to obtain tokens; then, the tokens are mapped to multi-dimensional vectors; at the same time, a position information vector is generated according to the formal position of the tokens in the text; finally, the two are superimposed as the final embedded code.

[0098] In the present invention, the argument fact tree derived from the preceding model is combined to retain the text position information in the process of generating the position information vector, and also increase the logical hierarchical relationship in the derivation process.

[0099] Step S5, embedding and encoding each of the embeddings having a hierarchical cascade relationship, sorting them by level, and inputting them into a transformer module, and the transformer module outputs an analysis result for the target problem in the photovoltaic business field.

[0100] In this step, combine Figure 2 , the transformer module includes an encoder, a decoder, a linear layer and a Softmax layer;

[0101] Embedding the embedding codes in a hierarchical cascade relationship, sorting them by level to form an embedding code sequence, and inputting the embedding code sequence into the encoder;

[0102] The encoder captures the global dependencies between the elements in the embedding coding sequence input, generates a series of context-aware representations, and inputs them to the decoder;

[0103] The decoder is used to generate an output sequence based on the context-aware representation; further, when generating the output sequence, the decoder considers both the context of the input sequence and the already generated output sequence.

[0104] The linear layer is used to map the output sequence of the decoder to a vector;

[0105] The Softmax layer is used to convert the vector output by the linear layer into a probability distribution.

[0106] This enables reasoning and analysis of photovoltaic business based on the argument fact tree.

[0107] The method provided by the present invention for improving the analysis effect of a large model in the photovoltaic business field based on the FHI logic deduction framework has the following advantages:

[0108] 1. The present invention realizes a vertical domain model solution in the photovoltaic field from semantic understanding to logical deduction and analysis combined with business facts;

[0109] 2. The present invention relies on the logical reasoning framework of Fact-Rational Hypothesis-System Investigation (FHI) to obtain the argument fact tree through the cyclic scheduling of semantic understanding and model agent;

[0110] 3. Use the hierarchical structure information of the argument fact tree to enhance the logical relationship logic in the position encoding, improve the depth of understanding of the transformer-type model on specific photovoltaic business issues, and realize reasoning and analysis.

[0111] The present invention realizes the model agent from semantic understanding to actual analysis and diagnosis, makes the model reasoning process conform to business logic, enhances the transformer model's understanding depth of photovoltaic business, and thus improves the accuracy and reliability of the prediction and analysis results of the large language model.

[0112] Combine the following Figure 3 and Figure 4 , taking the specific analysis process of "equipment health status of photovoltaic inverter at station A" as an example, a specific implementation example is introduced:

[0113] In this embodiment:

[0114] The basic root node is the Fact node "Equipment health status of photovoltaic inverter at station A";

[0115] By calling the FHI logical reasoning basic model once, the following logical reasoning tree is generated:

[0116] Relying on the "Large Model Efficiency Improvement Method Based on Knowledge Graph and Vector Retrieval Enhancement", four Hypothesize rational hypothesis nodes are generated, namely [H1-1] Which inverters are in Station A, [H1-2] Natural environmental factors of Station A, [H1-3] Equipment alarm information of Station A, and [H1-4] Which stations in the power grid are connected to Station A;

[0117]

H1-1

I1-1

F1-1

[0118]

H1-2

I1-2

F1-2

[0119]

H1-3

I1-3

[0120] [F1-3] The equipment alarm information of station A is as follows: 10kV bus voltage out-of-bounds alarm on July 15, 24; public connection point unbalance and flicker alarm; disconnector reset abnormality;"

[0121]

H1-4

I1-4

F1-4

[0122] Based on business logic, the above Fact sub-nodes are analyzed. Fact sub-nodes [F1-2] and [F1-3] are directly alarm information, and no longer call the FHI logic reasoning basic model in a loop. However, Fact sub-nodes [F1-1] and [F1-4] need to call the FHI logic reasoning basic model in a loop.

[0123] Therefore, after a generation of FHI derivation, the Fact subnode [F1-1] finally obtains the Fact subnode "[F1-1-1] B01 batch quality information, operating life 2.5 years, cumulative failure rate 5%, failure rate in the past year 1%;", and the Fact subnode "[F1-1-2] B02 batch quality information, operating life 4.5 years, cumulative failure rate 35%, failure rate in the past year 15%;"

[0124] After a generation of FHI derivation, the Fact subnode [F1-4] finally obtains the Fact subnode "[F1-4-1] Equipment alarm information of Station A: Public connection point imbalance alarm on July 15, 24;"

[0125] Based on business logic analysis, the Fact sub-nodes [F1-1-1], [F1-1-2], and [F1-4-1] directly reflect the health status of the equipment. Therefore, the FHI logical reasoning basic model is no longer called in a loop.

[0126] Finally get Figure 3 The FHI logical reasoning model shown; Figure 3 The FHI logical inference model shown is transformed into Figure 4 The argument fact tree shown.

[0127] By utilizing the hierarchical structure information of the argument fact tree, the logical connection of the position encoding is enhanced, the transformer-type model's understanding of specific photovoltaic business issues is improved, and reasoning and analysis are achieved.

[0128] In this solution, the argument fact tree derived from the previous model is combined to retain the text information position information in the process of generating the position information vector, and also increase the logical hierarchical relationship in the derivation process. The specific process is as follows: each word segment in the text is converted into a digital representation of semantics, and an M-dimensional text semantic vector is obtained, recorded as v_semantic; according to the text position information, an M-dimensional text position vector is obtained by combining sine and cosine with different periods, recorded as v_position; relying on the argument fact tree, according to each node level and the order within the layer, an initial hierarchical logical vector of dimension D is constructed, and then a structural position vector of dimension M is obtained through linear layer mapping, recorded as v_structure; finally, these three M-dimensional vectors are added to obtain an embedded code. In this embodiment, M is 512 and D is 2;

[0129] After that, it passes through the encoder (5 layers), decoder (5 layers), linear layer, and Softmax layer in turn, and finally outputs the probability distribution. Finally, the reasoning and analysis of the photovoltaic business based on the argument fact tree is realized.

[0130] In this embodiment, for the "equipment health status of the photovoltaic inverter at station A", the model outputs the following diagnostic results:

[0131] "Based on the information provided, the following inferences are made to assess the health status of the photovoltaic inverter equipment at station A:

[0132] 1. **Equipment age and failure rate**:

[0133] The inverters in batch B01 (NB001 and NB002) have been in operation for 2.5 years, with a cumulative failure rate of 5% and a failure rate of 1% in the past year, indicating that these devices are relatively new and in good overall operating condition with a low failure rate.

[0134] The inverters in batch B02 (NB003) have been in operation for 4.5 years, with a cumulative failure rate of 35%. The failure rate in the past year was 15%, indicating that these devices are older and have a significantly higher failure rate than batch B01, and there may be aging or design defects.

[0135] 2.**Environmental factors**:

[0136] Since Station A is located in a monsoon climate zone and there were heavy rain and high temperature warnings in July, these extreme weather conditions may have put additional pressure on the inverter equipment and increased the risk of failure.

[0137] 3. **Equipment alarm information**:

[0138] The voltage out-of-bounds alarm on July 15, 2024, may indicate that the inverter has problems coping with grid fluctuations, or that there is a problem with the grid itself.

[0139] Common connection point unbalance and flicker alarms may mean that the inverter has problems with output power quality, which may be caused by a fault in the inverter itself or a problem in the external grid.

[0140] Abnormal reset of the isolating switch may indicate a problem with the maintenance or operation of the equipment and requires further inspection.

[0141] 4.**Grid ownership and group impact**:

[0142] Station A and Station A2 belong to the same power generation group, and Station A2 also has an imbalance alarm, which may indicate that there are problems with the grid stability within the group, affecting the normal operation of the inverter.

[0143] Based on the above information, the following conclusions can be drawn:

[0144] **B01 batch inverters (NB001 and NB002)**: Currently in good health, but need to continue to be monitored, especially in extreme weather conditions, to prevent potential failures.

[0145] **B02 batch inverter (NB003)**: Poor health status, high failure rate, may require immediate detailed inspection and maintenance, or consider replacement in advance to avoid possible long downtime.

[0146] **Overall grid status**: There may be stability issues and coordination with the grid operator is required to check the grid status within the group to reduce the impact on the inverters.

[0147] **Maintenance and Monitoring**: It is recommended to strengthen the daily maintenance and monitoring of equipment, especially during extreme weather events, as well as timely response and processing of alarm information.

[0148] **Preventive Measures**: Taking into account environmental factors and the age of the equipment, a preventive maintenance plan should be developed, including regular inspection and replacement of wearing parts, to extend the life of the equipment and keep it running efficiently.

[0149] "

[0150] The present invention provides a method for improving the analysis effect of a large model in the photovoltaic business field based on the FHI logic deduction framework, so that the large language model in the photovoltaic business field not only understands photovoltaic professional knowledge, but also can realize long-link reasoning analysis based on the reasoning logic chain and the business interface of the scheduling information system, so that the reasoning process of the large language model conforms to the business logic, and enhances the depth of understanding of the photovoltaic business by the large language model, thereby improving the accuracy and reliability of the prediction and analysis results of the large language model.

[0151] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be considered as the scope of protection of the present invention.

Claims

1. A method for improving the analysis effect of large models in the photovoltaic business field based on the FHI logic deduction framework, characterized in that: The following steps are involved: Step S1, establishing a basic model of FHI logic reasoning; the basic model of FHI logic reasoning is a tree structure, including a Fact node, a Hypothesize node, an Investigate node and a Fact sub-node connected hierarchically from top to bottom; Step S2, for the target problem in the photovoltaic business field, repeatedly call the FHI logical reasoning basic model to obtain the FHI logical reasoning model; the method is: After the FHI logic reasoning basic model matching the target problem in the photovoltaic business field is established, the associated business knowledge and business logic are called, and each of the Fact fact sub-nodes is analyzed to determine whether the Fact fact sub-node is the bottom-level node associated with the target problem in the photovoltaic business field in the business logic. If so, the re-calling of the FHI logic reasoning basic model is terminated; if not, the Fact fact sub-node is used as the Fact fact node, and the FHI logic reasoning basic model is called again to increase the level of the tree structure. If the cycle is continuous, the FHI logic reasoning model is finally established; Step S3, converting the FHI logical reasoning model into an argument fact tree; the argument fact tree only has Fact fact nodes and Fact fact sub-nodes of each level; Step S4, analyzing and processing each of the Fact sub-nodes in the argument fact tree to obtain an embedding code embedded with the hierarchical features in the argument fact tree; Step S5, embedding and encoding each of the embeddings having a hierarchical cascade relationship, sorting them by level, and inputting them into a transformer module, and the transformer module outputs an analysis result for the target problem in the photovoltaic business field.

2. According to claim 1, a method for improving the analysis effect of a large model in the photovoltaic business field based on the FHI logic deduction framework is characterized in that: In step S1, the process of establishing the FHI logical reasoning basic model is as follows: Step S1.1, storing the semantic description information of the target problem into the Fact node; Step S1.2, calling the business knowledge and business logic associated with the target problem, analyzing and reasoning the semantic description information of the target problem stored in the Fact node, obtaining N rational hypothesis texts associated with the target problem, and for each rational hypothesis text, generating a Hypothesize rational hypothesis node as a child node of the Fact node; Step S1.3, for each of the Hypothesize rational hypothesis nodes, obtain M associated Investigate system investigation nodes and use them as child nodes of the Hypothesize rational hypothesis node; each of the Investigate system investigation nodes is bound to a model agent, and by executing the model agent, text-based argument facts are obtained; Step S1.4, the text-type argument facts of each Investigate system investigation node are stored in the Fact fact child node which is a child node of the Investigate system investigation node; From top to bottom, the FHI logical reasoning basic model is established to obtain a tree structure formed by Fact nodes, Hypothesize nodes, Investigate nodes and Fact sub-nodes.

3. According to claim 2, a method for improving the analysis effect of a large model in the photovoltaic business field based on the FHI logic deduction framework is characterized in that: The business knowledge and business logic are a large language model enhanced based on knowledge graph and vector retrieval.

4. According to claim 2, a method for improving the analysis effect of a large model in the photovoltaic business field based on the FHI logic deduction framework is characterized in that: The model agent bound to the Investigate system investigation node is a model agent obtained by training with historical data; The training method is: The model agent is trained using a key-value pair in the form of {"text":(model agent name, key parameter list)}, so that the model agent has the capabilities of semantic understanding, top_N model agent recall, and key parameter extraction.

5. According to claim 4, a method for improving the analysis effect of a large model in the photovoltaic business field based on the FHI logic deduction framework is characterized in that: The model agent comprises an input processing submodule, a business interface calling submodule and an integrated output submodule; The input processing submodule is used to perform semantic understanding on the rational hypothesis text input from the Hypothesize rational hypothesis node; The business interface calling submodule is used to call the associated business database or business data table according to the semantic understanding of the input processing submodule, and query and obtain key parameters associated with the semantic understanding as query results; The integration output submodule is used to form a key parameter table with the multiple key parameters obtained by the business interface calling submodule, and output it to the Fact fact subnode.

6. According to claim 1, a method for improving the analysis effect of a large model in the photovoltaic business field based on the FHI logic deduction framework is characterized in that: In step S3, the FHI logical reasoning model is converted into an argument fact tree, specifically: Delete each Hypothesize rational hypothesis node and Investigate system investigation node in the FHI logical reasoning model, keep the number of levels and cascade relationship of each Fact fact sub-node unchanged, and obtain the argument fact tree; wherein the number of levels and cascade relationship are called argument derivation level relationship.

7. A method for improving the analysis effect of a large model in the photovoltaic business field based on the FHI logic deduction framework according to claim 6, characterized in that: Step S4 is specifically as follows: The Fact sub-node stores text-based argument facts and argument derivation hierarchical relationships; Segmenting the text-like argument facts, wherein the text-like argument facts have M-dimensional segmentations, and each segmentation is converted into a digital representation of semantics, thereby obtaining an M-dimensional text semantic vector v_semantic; According to the text position information of each word in the text class argument fact, an M-dimensional text position vector v_position is obtained by combining sine and cosine with different periods; Relying on the argument fact tree, the hierarchical relationship is deduced according to the arguments of the Fact fact sub-node, including the hierarchical level and the order within the level, to form an initial hierarchical logic vector with a dimension of D, and then a structural position vector v_structure with a dimension of M is obtained through linear layer mapping; The text semantic vector v_semantic, the text position vector v_position and the structure position vector v_structure are added together to obtain the embedding encoding embeding of the text class argument fact.

8. The method for improving the analysis effect of large models in the photovoltaic business field based on the FHI logic deduction framework according to claim 1 is characterized in that: Step S5 is specifically as follows: The transformer module includes an encoder, a decoder, a linear layer and a Softmax layer; Embedding the embedding codes in a hierarchical cascade relationship, sorting them by level to form an embedding code sequence, and inputting the embedding code sequence into the encoder; The encoder captures the global dependencies between the elements in the embedding coding sequence input, generates a series of context-aware representations, and inputs them to the decoder; The decoder is used to generate an output sequence based on the context-aware representation; The linear layer is used to map the output sequence of the decoder to a vector; The Softmax layer is used to convert the vector output by the linear layer into a probability distribution.

Images