Analysis management unit, and interactive management system and method for photovoltaic power station
By designing an interactive management system for photovoltaic power stations and using large language models to process users' natural language instructions, it solves the problem that ordinary users have difficulty in monitoring and maintaining autonomously, and realizes efficient management and user-friendly experience of photovoltaic power stations.
Patent Information
- Application Number
- PCT/CN2024/101761
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2024-04-29
- Filing Date
- 2024-06-26
- Publication Date
- 2024-11-07
AI Technical Summary
Traditional photovoltaic power station management systems are highly professional, and it is difficult for ordinary users to monitor and maintain independently, and lack a user-friendly interactive management system.
An interactive management system for photovoltaic power stations is designed, including terminals, equipment terminals and clouds. The analysis management unit uses a large language model to process natural language instructions input by users, generate feedback instructions or control instructions, simplify user input and lower user thresholds. And improve management efficiency.
It realizes all-weather and comprehensive management of photovoltaic power stations, reduces user input requirements, improves user experience, enhances the system's equipment control and processing capabilities, and can better respond to the management needs of photovoltaic power stations.
Smart Images

Figure CN2024101761_07112024_PF_FP_ABST
Abstract
Description
An analysis management unit, photovoltaic power station interactive management system, and method Technical Field
[0001] The present invention relates to the field of photovoltaic control technology, in particular to the field of photovoltaic power station management technology, and specifically to an analysis and management unit, a photovoltaic power station interactive management system, and a method. Background Art
[0002] A photovoltaic power station is a power generation system that uses solar energy and adopts special materials such as crystalline silicon panels, inverters and other electronic components. It is connected to the power grid and transmits electricity to the grid. Photovoltaic power stations can be divided into independent power generation systems with batteries and grid-connected power generation systems without batteries.
[0003] Photovoltaic power station management system is an operation and maintenance system used in photovoltaic power stations. It plays a very important role in maintaining the normal operation of photovoltaic power stations. However, traditional photovoltaic power station management systems require high professionalism and high requirements for operation and maintenance personnel, not to mention ordinary users. Ordinary users are basically unable to conduct independent monitoring and maintenance of photovoltaic power stations.
[0004] Therefore, developing a system and method that can manage photovoltaic power plants facing users is a technical problem that needs to be solved urgently.
[0005] Summary of the Invention
[0006] A main purpose of the present invention is to overcome at least one of the above-mentioned drawbacks, and to provide an analysis management unit to improve the efficiency of interaction with users and enhance user experience.
[0007] A main purpose of the present invention is to overcome at least one of the above-mentioned drawbacks, and to provide a photovoltaic power station interactive management system and a photovoltaic power station interactive management method that are highly interactive and user-friendly.
[0008] In order to achieve the above object, the technical solution adopted by the present invention is:
[0009] The present invention provides an interactive management system for a photovoltaic power station, which includes:
[0010] A terminal equipped with an interactive APP, wherein the terminal receives instructions input by a user;
[0011] The equipment end is installed at the photovoltaic power station, and the equipment end controls and manages the photovoltaic power station;
[0012] The cloud is connected to the terminal and the device;
[0013] An analysis and management unit, configured on the terminal, device, or cloud, receives the instruction and performs analysis and processing to form an output instruction, which, after parsing, forms a feedback instruction or an executable control instruction;
[0014] The control instructions are distributed to the terminal, cloud or device to perform corresponding execution actions.
[0015] According to one embodiment of the present invention, the analysis management unit includes:
[0016] An input preprocessing module is used to receive the instruction and perform feature capture on the instruction to form valid feature data in text form;
[0017] A context comprehensive processing module is connected to the input preprocessing module, receives the valid feature data, and retrieves and calls prompt words that match the valid feature data to form a comprehensive input instruction;
[0018] The large language model has an input side connected to the input preprocessing module and an output side connected to the output parsing module, receives the comprehensive input instructions and processes them to form output instructions, and the output instructions are output instructions based on natural language.
[0019] According to one embodiment of the present invention, the analysis and management unit further includes a memory storage module, which is used to store photovoltaic standard data and real-time interaction data. The memory storage module is connected to the context integration processing module.
[0020] According to one embodiment of the present invention, the context comprehensive processing module extracts the data stored in the memory storage module based on the valid feature data to form a first input instruction, and the first input instruction is a prompt word prompt. The first input instruction is fed back to the input preprocessing module, and the input preprocessing module aggregates the valid feature data and the first input instruction to form a comprehensive input instruction.
[0021] According to one embodiment of the present invention, the analysis management unit further includes an output parsing module, the output side of the large language model is connected to the output parsing module, and the output parsing module receives the output instruction and parses the output instruction to form the feedback instruction and / or the control instruction.
[0022] According to one embodiment of the present invention, the memory storage module includes a permanent storage submodule and a flash memory submodule, the permanent storage submodule is used to store photovoltaic standard data, and the flash memory submodule is used to store real-time interactive data.
[0023] According to one embodiment of the present invention, the photovoltaic standard data includes API call documents, product descriptions, and photovoltaic industry technical information.
[0024] According to one embodiment of the present invention, the large language model includes a ChatGPT model.
[0025] According to one embodiment of the present invention, the establishment of the large language model includes the following training process:
[0026] Acquire photovoltaic-related basic data and instruction data to construct a domain data set, and annotate the domain data set to obtain an annotated domain data set;
[0027] Training the pre-trained language model using the domain dataset to update the pre-trained language model, and training the updated pre-trained language model using the annotated domain dataset to obtain a supervised fine-tuning model;
[0028] The supervised fine-tuning model adds a classification mechanism, and then combines manual feedback to score and rank the output of the supervised fine-tuning model after adding classification, forming a reward model with a scoring mechanism;
[0029] The domain dataset is used as input to train the reinforcement learning model, the obtained output result is sent to the reward model for scoring, and the obtained answer score is fed back to the reinforcement learning model for adjustment and reinforcement of the reinforcement learning model;
[0030] The obtained adjusted and reinforced reinforcement learning model is the final large language model.
[0031] According to one embodiment of the present invention, when the domain dataset trains the pre-trained language model, part of the parameter layers of the pre-trained language model are frozen, and during training, the domain dataset only updates the parameters of another part of the parameter layers in the pre-trained language model.
[0032] According to one embodiment of the present invention, when the annotated domain dataset is used to train the updated pre-trained language model, the model data is updated by a back-propagation algorithm to obtain the supervised fine-tuning model.
[0033] According to one embodiment of the present invention, when the reinforcement learning model is adjusted and reinforced, a KL divergence constraint with the supervised fine-tuning model is added to the loss function of the reinforcement learning model to prevent the reinforcement learning model from overfitting.
[0034] According to one embodiment of the present invention, the photovoltaic-related basic data includes one or more of the user manual of the photovoltaic power station, frequently asked questions and answers, technical documents, online materials, and user feedback information, and the instruction data is obtained by parsing the system instruction set and interface parameters of the photovoltaic power station.
[0035] According to one embodiment of the present invention, the system further includes a routing unit, which is connected to the analysis and management unit and is used to distribute the control instructions to the terminal, cloud or device end for performing corresponding execution actions.
[0036] According to one embodiment of the present invention, the routing unit is configured in the cloud.
[0037] According to one embodiment of the present invention, the system further includes a plurality of call execution units, each of which is connected to the routing unit, and the routing unit distributes control instructions to the call execution unit at the terminal, cloud or device end to perform the corresponding execution action.
[0038] According to one embodiment of the present invention, the calling execution unit is respectively configured at the terminal, the cloud or the device.
[0039] According to one embodiment of the present invention, the system further includes an action response unit, which is used to monitor the execution status of the action executed at the terminal, cloud or device.
[0040] According to one embodiment of the present invention, the input end of the action response unit is connected to each calling execution unit used to perform the execution action, the output end of the action response unit is connected to the input preprocessing module, and the calling execution unit determines to call the terminal, cloud or device end to perform the corresponding execution action according to the control instruction. During the execution of the action, the action response unit feeds back response status data containing the execution status to the context comprehensive processing module, and the execution action is recursively performed as needed.
[0041] According to one embodiment of the present invention, the action response unit is configured in the terminal.
[0042] According to one embodiment of the present invention, the memory storage module includes a storage management submodule, which is connected to the action response unit. The storage management submodule receives the response status data fed back by the action response unit and updates it to the real-time interaction data.
[0043] According to one embodiment of the present invention, the real-time interaction data includes effective feature data and response status data.
[0044] According to one embodiment of the present invention, it also includes an identity setting unit, which is connected to the context integration processing module. The identity setting unit sets the working scenario and role of the large language model according to external input or the valid feature data.
[0045] According to one embodiment of the present invention, it further includes a user information management unit, which is connected to the context integration processing module, and stores the user's account information and the device information bound to the account.
[0046] According to one embodiment of the present invention, the instruction input by the user is voice or text input.
[0047] According to one embodiment of the present invention, the analysis management unit receives the instruction and performs analysis and processing to form an output instruction based on natural language.
[0048] In particular, the present invention also provides a photovoltaic power station interactive management method, which includes the following steps:
[0049] The user inputs instructions through the interactive APP at the terminal, and sends the input instructions to the analysis management unit;
[0050] The instruction is analyzed and processed by the analysis and management unit to form an output instruction, and the output instruction is then formatted and parsed to form a feedback instruction or an executable control instruction. The analysis and management unit is configured in the terminal, device or cloud;
[0051] For the feedback instruction, it is sent back to the interactive APP to form interactive communication with the user;
[0052] The executable control instructions are in a machine language format after formatting and parsing, and are distributed to the terminal, the cloud or the equipment installed at the photovoltaic power station according to the specific content of the control instructions to perform the corresponding execution actions.
[0053] According to one embodiment of the present invention, the process of analyzing and processing the instruction at the analysis management unit to form an output instruction includes:
[0054] Preprocessing the instruction to generate effective feature data in text form by capturing features of keywords in the instruction;
[0055] Searching for data from a memory storage module according to the valid feature data to form a first input instruction, wherein the first input instruction is a prompt word, and combining the first input instruction with the valid feature data to form a comprehensive input instruction;
[0056] The comprehensive input instruction is read and processed by a large language model to form an output instruction.
[0057] According to one embodiment of the present invention, the process of searching for data from a memory storage module according to valid feature data to form a first input instruction includes:
[0058] According to the effective characteristic data, data is searched from the memory storage module to search for photovoltaic standard data and / or real-time interactive data corresponding to the effective characteristic data, and the found data is aggregated to form the first input instruction.
[0059] According to one embodiment of the present invention, the output instruction is further output parsed to form the feedback instruction and / or the control instruction, and the control instruction is in a machine language format after formatting and parsing.
[0060] According to one embodiment of the present invention, the process of pre-processing the instruction includes a step of cleaning the input instruction.
[0061] According to one embodiment of the present invention, a support vector search method is used in the process of searching for photovoltaic standard data and / or real-time interactive data corresponding to the effective characteristic data.
[0062] According to one embodiment of the present invention, the feedback instruction is sent back to the interactive communication process between the interactive APP and the user.
[0063] The new input instructions of the user are combined with the feedback instructions to form new effective feature data, and then the data search process is performed to form a new comprehensive input instruction which is sent to the large language model for processing.
[0064] According to one embodiment of the present invention, the executable control instructions are distributed to the terminal, the cloud or the device installed at the photovoltaic power station according to the specific content of the control instructions, and the corresponding execution actions are performed.
[0065] For an execution action with multiple subtasks, the execution action is disassembled, and the system recursively executes each subtask automatically.
[0066] According to one embodiment of the present invention, after the analysis and processing at the analysis management unit generates valid feature data, the instruction automatically sets the working scenario and role of the large language model according to the valid feature data.
[0067] According to one embodiment of the present invention, when a user inputs instructions through an interactive APP at a terminal, the user manually inputs identity setting parameters to manually set the working scenario and role of the large language model.
[0068] According to one embodiment of the present invention, when the analysis management unit searches for data from the memory storage module based on valid feature data, it simultaneously retrieves the user's account information and the device information bound to the account, and summarizes the account information and device information together in the first input instruction.
[0069] In particular, the present invention further provides an analysis management unit, wherein the analysis management unit forms interactive communication with the user through an interactive APP, the analysis management unit receives the instructions input by the user and performs analysis and processing to form output instructions, and the output instructions form feedback instructions or executable control instructions after parsing, wherein the feedback instructions are sent back to the interactive APP to form interactive communication with the user, and the control instructions are distributed to the terminal, cloud or device end to perform corresponding execution actions;
[0070] The analysis management unit includes:
[0071] An input preprocessing module is used to receive the instruction and capture features of the instruction to form valid feature data;
[0072] A context synthesis processing module is connected to the input pre-processing module, receives the valid feature data, and retrieves and calls data matching the valid feature data to form a comprehensive input instruction;
[0073] The large language model, whose input side is connected to the input preprocessing module, receives the comprehensive input instructions and processes them to form output instructions.
[0074] According to one embodiment of the present invention, the analysis and management unit also includes a memory storage module, which is used to store photovoltaic standard data and real-time interaction data. The memory storage module is connected to the context comprehensive processing module, wherein the photovoltaic standard data includes API call documents, product descriptions, and photovoltaic industry technical information.
[0075] According to one embodiment of the present invention, the memory storage module includes a permanent storage submodule and a flash memory submodule, the permanent storage submodule is used to store photovoltaic standard data, and the flash memory submodule is used to store real-time interactive data.
[0076] According to one embodiment of the present invention, the context comprehensive processing module extracts the data stored in the memory storage module based on the valid feature data to form a first input instruction, and feeds the first input instruction back to the input preprocessing module, in which the valid feature data and the first input instruction are aggregated to form a comprehensive input instruction.
[0077] According to one embodiment of the present invention, the analysis management unit further includes an output parsing module, the output side of the large language model is connected to the output parsing module, and the output parsing module receives the output instruction and parses the output instruction to form the feedback instruction and / or the control instruction.
[0078] According to one embodiment of the present invention, the establishment of the large language model includes the following training process:
[0079] Acquire photovoltaic-related basic data and instruction data to construct a domain data set, and annotate the domain data set to obtain an annotated domain data set;
[0080] Training the pre-trained language model using the domain dataset to update the pre-trained language model, and training the updated pre-trained language model using the annotated domain dataset to obtain a supervised fine-tuning model;
[0081] The supervised fine-tuning model adds a classification mechanism, and then combines manual feedback to score and rank the output of the supervised fine-tuning model after adding classification, forming a reward model with a scoring mechanism;
[0082] The domain dataset is used as input to train the reinforcement learning model, the obtained output result is sent to the reward model for scoring, and the obtained answer score is fed back to the reinforcement learning model for adjustment and reinforcement of the reinforcement learning model;
[0083] The obtained adjusted and reinforced reinforcement learning model is the final large language model.
[0084] According to one embodiment of the present invention, when the domain dataset trains the pre-trained language model, part of the parameter layers of the pre-trained language model are frozen, and during training, the domain dataset only updates the parameters of another part of the parameter layers in the pre-trained language model.
[0085] According to one embodiment of the present invention, when the annotated domain dataset is used to train the updated pre-trained language model, the model data is updated by a back-propagation algorithm to obtain the supervised fine-tuning model.
[0086] According to one embodiment of the present invention, when the reinforcement learning model is adjusted and reinforced, a KL divergence constraint with the supervised fine-tuning model is added to the loss function of the reinforcement learning model to prevent the reinforcement learning model from overfitting.
[0087] According to one embodiment of the present invention, the photovoltaic-related basic data includes one or more of the user manual of the photovoltaic power station, frequently asked questions and answers, technical documents, online materials, and user feedback information, and the instruction data is obtained by parsing the system instruction set and interface parameters of the photovoltaic power station.
[0088] In particular, the present invention also provides a computing device, comprising a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor implements the method described above when executing the computer program.
[0089] The present invention also provides a computer-readable storage medium having a computer program stored therein, wherein the computer program implements the above method when executed by a processor. The computer-readable storage medium is preferably a non-volatile readable storage medium.
[0090] Compared with the prior art, the advantages and beneficial effects of the photovoltaic power station interactive management system and the photovoltaic power station interactive management method of the present invention patent application are:
[0091] The photovoltaic power station interactive management system of the present application simplifies the user input at the terminal and greatly reduces the input requirements, thereby lowering the user threshold for the photovoltaic power station interactive management system. While improving the user experience, it can better realize all-weather and all-round management of the photovoltaic power station.
[0092] Furthermore, this application innovatively introduces a large language model (such as the ChatGPT model). The input side of the large language model uses comprehensive input instructions in text form, and the output side uses output instructions based on natural language. From front-end input to back-end processing output, it is all based on natural language processing, which is more in line with the usage needs of traditional users. During the final specific execution, the output instructions based on natural language are converted into a format and then used for specific execution, which reduces the processing difficulty of this management method, and the data sources in the processing process are more diverse, making the system processing capability stronger and the method more comprehensive in dealing with the problems to be processed. BRIEF DESCRIPTION OF THE DRAWINGS
[0093] Hereinafter, some specific embodiments of the present invention will be described in detail in an exemplary and non-limiting manner with reference to the accompanying drawings. The same reference numerals in the accompanying drawings indicate the same or similar components or parts. It should be understood by those skilled in the art that these drawings are not necessarily drawn to scale. In the accompanying drawings:
[0094] FIG1 is a structural block diagram of an interactive management system for a photovoltaic power station according to an embodiment of the present invention;
[0095] FIG2 is a structural diagram of a computing device according to another embodiment of the present application;
[0096] FIG3 is a structural diagram of a computer-readable storage medium according to another embodiment of the present application;
[0097] FIG4 is a schematic diagram of a process for establishing a large language model according to an embodiment of the present invention;
[0098] Figure 5 is a schematic diagram comparing the graphics card memory usage of the traditional training method and the graphics card memory usage after using ZeRO graphics memory optimization technology;
[0099] FIG6 is a schematic diagram of a workflow of a photovoltaic power station startup method according to one embodiment of the present application;
[0100] FIG7 is an interface for performing photovoltaic knowledge question and answering in an interactive management system according to one embodiment of the present application;
[0101] FIG8 is an interface of an interactive management system performing multimodal question and answer according to one embodiment of the present application;
[0102] FIG9 is an interface for performing questions and answers on real-time information of a photovoltaic power station through an interactive management system according to one embodiment of the present application;
[0103] FIG10 is an interface of an interactive management system for performing mode configuration after receiving an instruction according to one embodiment of the present application;
[0104] FIG11 is a diagrammatic display interface of an interactive management system according to one embodiment of the present application;
[0105] FIG12 is a schematic diagram of an interface effect of an interactive APP according to an embodiment of the present invention. DETAILED DESCRIPTION
[0106] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.
[0107] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0108] Example 1:
[0109] This embodiment describes an interactive management system for a photovoltaic power station, as shown in FIG1 , which includes:
[0110] A terminal equipped with an interactive APP, wherein the terminal receives instructions input by a user;
[0111] A cloud end is in communication with the terminal and the device end. In this embodiment, an analysis and management unit is configured in the cloud end. The analysis and management unit receives the instruction and performs analysis and processing to form an output instruction based on natural language. The output instruction is parsed to form a feedback instruction or an executable control instruction;
[0112] The device side is installed at the photovoltaic power station and controls and manages the photovoltaic power station. A photovoltaic power station generally consists of a solar cell array, a battery pack, a charge and discharge controller, an inverter, an AC power distribution cabinet, a solar tracking control system, and other equipment. In this embodiment, the device side refers to the aforementioned charge and discharge controller, inverter, AC power distribution cabinet, solar tracking control system, and other equipment.
[0113] The control instructions are distributed to the terminal, cloud or device to perform corresponding execution actions.
[0114] Furthermore, unlike the present embodiment where the analysis and management unit is located in the cloud, the analysis and management unit can also be configured on the terminal or device side, and can be adjusted based on the user's usage requirements and the data processing performance of each side. It is even technically possible to rationally split the functional modules in the analysis and management unit and place them in one of the three: the terminal, the device side, or the cloud side. In other words, the processing of the analysis and management unit can be implemented by combining the functional modules of the terminal, the device side, or the cloud side.
[0115] The instructions input by the user at the terminal are voice or text input. It is not difficult to find that in this embodiment, the initial user input is natural language input, and the output instructions formed by the analysis and management unit are also based on natural language. When the user "communicates" with the terminal in a regular manner (not involving specific execution control instructions), it is all based on natural language. As shown in the interface effect of the interactive APP in Figure 12, the user can completely input instructions through natural language, but can interact with the background in a precise and professional manner, making communication smoother and more convenient.
[0116] In other words, the interactive management system for photovoltaic power plants in this embodiment simplifies user input at the terminal, significantly reducing input requirements. This lowers the user threshold for the interactive management system, enabling guided management operations without the need for traditional technicians. This improves user experience while facilitating data interaction with the cloud and devices anytime, anywhere, enabling better all-weather, comprehensive management of photovoltaic power plants.
[0117] Taking the start-up of a photovoltaic power station and user voice input as an example, as shown in Figure 6, the user (power station owner or technician) wakes up the interactive app at the terminal through voice interaction. The user tells the interactive app "I want to start the station" through voice interaction. The interactive app calls the relevant data for the start-up from the cloud and fills in the default data. The user can use voice interaction on the interactive app to modify the parameters of the photovoltaic power station one by one. Based on the modified parameters, the application program interface of the device side (such as the photovoltaic inverter) is called to complete the setting or modification of the corresponding parameters. In other words, the interactive app can automatically perform a series of start-up operations according to the user's voice input requirements. If there are any problems during the process, the interactive app can also inform the user and provide solutions through voice broadcast or text dialog box.
[0118] It is understandable that it is preferred that the instructions input by the user are voice, as the form of voice dialogue is more flexible and can give the user a better user experience.
[0119] In one embodiment, the analysis management unit includes:
[0120] An input preprocessing module is used to receive the instruction and capture features of the instruction to form valid feature data;
[0121] A context comprehensive processing module is connected to the input preprocessing module, receives the valid feature data, and retrieves and calls prompt words that match the valid feature data to form a comprehensive input instruction;
[0122] The large language model, whose input side is connected to the input preprocessing module and whose output side is connected to the output parsing module, receives the comprehensive input instructions and processes them to form output instructions based on natural language.
[0123] The input preprocessing module cleans and filters user input commands, as it's important to understand that some user input commands are invalid or non-compliant, not related to the specific operation and management of the photovoltaic power plant. These commands need to be filtered out and the user prompted to re-enter the command until authentic, legal, and valid feature data can be generated. The context-integrated processing module, as its name suggests, bridges the gap between the previous (introducing valid feature data and the foundation for other previous operations) and the next (searching and retrieving corresponding data for the newly introduced valid feature data), thereby merging and constructing the comprehensive input commands.
[0124] Since the effective feature data itself is a language feature, that is, text, and the data searched and retrieved also come from API call documents, product descriptions, photovoltaic industry technical information, etc., and are therefore also composed of text, the resulting comprehensive input instructions are in the form of text. After being input into the large language model, the comprehensive input instructions are trained and integrated to form output instructions based on natural language. This output instruction can be converted into voice to interact with the user according to its processing requirements, which is the feedback instruction mentioned above. In this embodiment, it is a feedback instruction in the form of voice. This output instruction can also be parsed into a control instruction that can control the terminal, cloud or device end according to its processing requirements.
[0125] In this embodiment, the large language model used is the currently popular ChatGPT model. The ChatGPT model is a natural language processing tool driven by artificial intelligence technology. It can conduct conversations by understanding and learning human language, and can also interact according to the context of the chat. Therefore, the context comprehensive processing module of this embodiment outputs the comprehensive input instructions to the non-word single word model to form an output instruction based on natural language. This ChatGPT model can be trained using an open source algorithm to obtain a model or deployed after fine-tuning an open source model. Taking site opening as an example, fine-tune can convert the site opening process into a series of conversations, and then use these conversations as training data to fine-tune the large language model to make it more suitable for site opening scenarios.
[0126] This application uses photovoltaic field-related data to train the large language model to improve the model's understanding and application capabilities of professional knowledge in the photovoltaic field, and accurately and effectively embeds the optimized large language model into the photovoltaic management process to achieve intelligent interaction between users and the system, provide intelligent problem identification and solving functions, and enable users to ask questions or instructions in natural language, and the system can accurately understand and give corresponding answers or operation suggestions.
[0127] Existing technologies for answering questions in the photovoltaic field often rely on experts for offline or online Q&A. However, both online and offline approaches suffer from issues such as incomplete knowledge, delayed responses, and inaccurate answers. Directly applying large language models (LLMs) to real-world scenarios, such as photovoltaic Q&A systems, also encounters difficulties because LLMs cannot use external knowledge for reference. For example, for questions like "What does the alarm at power station A mean?", "What is the power output of inverter model B?", and "What is the power generation of power station C today?", traditional large language models may produce answers that are inconsistent with the actual situation. Furthermore, photovoltaic power station management systems also require device control capabilities to monitor, adjust, and manage photovoltaic equipment. However, existing technologies have limitations in this regard, such as a lack of intelligent device control strategies and an inability to respond to device status changes in real time.
[0128] The interactive APP in this application can display various work scenarios. As shown in Figure 7, it shows the interface of the interactive management system of this application for photovoltaic knowledge question and answer. For example, if the user asks "What is the power of PV?", the interactive management system can search for relevant information from the technical data of the photovoltaic industry and display it, and wait for further inquiries from the user; as shown in Figure 8, it shows the interface of the interactive management system of this application for multimodal question and answer, and the status of the indicator light is displayed in the figure; as shown in Figure 9, it shows the interface for question and answer of real-time information of photovoltaic power station, for example, according to the instruction input by the user "What is the power generation today", the power generation data of the photovoltaic power station today is quickly captured and displayed, and then compared. If the user inputs "What are today's alarm information?", the command is analyzed and processed, and the alarm information of the photovoltaic power station on that day is captured as a feedback command and displayed; as shown in Figure 10, the interface of the interactive management system of the present application for mode configuration after receiving the command is shown. This interface quickly switches the display mode of the interface according to the command "Set dark mode" or "Set light mode" input by the user; as shown in Figure 11, the interactive management system of the present application is shown in the form of a chart display interface, which can accurately reflect the current power generation status of the day in the form of a chart according to the command "Today's power generation curve" input by the user.
[0129] Therefore, current technologies for question-answering systems and device control in the photovoltaic field suffer from problems such as incomplete knowledge, inaccurate answers, and insufficient device control capabilities. To address these issues, this application proposes a method for fine-tuning and training a large language model based on specific domain knowledge, as well as a corresponding system. The goal is to improve the large language model's question-answering capabilities in the field of photovoltaic power plant management systems, optimize the system's interactive efficiency and user experience, and enhance the system's device control capabilities, enabling intelligent monitoring and management of photovoltaic equipment.
[0130] The interactive power plant management system proposed in this application can be applied to all aspects of the interactive management process for photovoltaic power plants, including equipment fault diagnosis, user consultation and answering, and data monitoring and analysis. By integrating large language models with the photovoltaic power plant management system, it is possible to rapidly implement and promote intelligent technologies in the photovoltaic sector, providing strong support for its intelligent development.
[0131] In one embodiment, as shown in FIG4 , the establishment of the large language model includes the following training process:
[0132] Acquire photovoltaic-related basic data and instruction data to construct a domain data set, and annotate the domain data set to obtain an annotated domain data set;
[0133] Training the pre-trained language model using the domain dataset to update the pre-trained language model, and training the updated pre-trained language model using the annotated domain dataset to obtain a supervised fine-tuning model;
[0134] The supervised fine-tuning model adds a classification mechanism, and then combines manual feedback to score and rank the output of the supervised fine-tuning model after adding classification, forming a reward model with a scoring mechanism;
[0135] The domain dataset is used as input to train the reinforcement learning model, the obtained output result is sent to the reward model for scoring, and the obtained answer score is fed back to the reinforcement learning model for adjustment and reinforcement of the reinforcement learning model;
[0136] The obtained adjusted and reinforced reinforcement learning model is the final large language model.
[0137] In this embodiment, the establishment of the domain dataset is data preparation, which includes data collection, data cleaning, data labeling, data processing and dataset division. Specifically:
[0138] Data Collection: First, determine the source of data collection. PV-related basic data may include, but is not limited to, PV system user manuals, FAQs, technical documentation, online resources (such as forum posts, professional websites, forums, and social media related to PV systems, which capture general domain knowledge and provide more real-time, specific Q&A), and user feedback. Command data is obtained by parsing the system command set and interface parameters of a specific PV power plant.
[0139] Take the common questions and answers as an example to construct the basic data. For example, the questions corresponding to the natural language input are used as the content, and the corresponding answers are used as the target. The following are some examples of content and corresponding targets:
[0140] Taking the system instruction set and interface parameters of a photovoltaic power station as an example, the corresponding system instructions are input in natural language as the content, and the corresponding interface parameters are used as the target. The following are some examples of content and corresponding targets:
[0141] Data cleaning: The collected basic data may contain a large amount of noise and invalid information, so data cleaning and preprocessing are required. Data cleaning includes steps such as removing duplicate data, correcting erroneous information, and filtering out irrelevant information to ensure data quality and accuracy.
[0142] Data labeling: After data cleaning, it needs to be labeled to accurately understand the meaning and relationships of the data during model training. Data labeling may include pairing questions and answers, classifying and labeling instruction data, etc., to ensure that the labeled domain dataset can provide effective supervision signals for the model.
[0143] Data processing: Depending on the specific training requirements of each model, the data may need to be further processed and converted. Such further processing and conversion may include operations such as word segmentation, vectorization, and serialization of text data to facilitate model input and processing.
[0144] Dataset Partitioning: Finally, the domain dataset needs to be divided into training, validation, and test sets. Similarly, the labeled domain dataset also needs to be divided into training, validation, and test sets for model training, evaluation, and testing. Typically, the training set is used for model training, the validation set is used to adjust the model's hyperparameters and structure, and the test set is used to ultimately evaluate the model's performance and generalization capabilities.
[0145] Through the above steps, data related to photovoltaic system questions and answers or instructions can be effectively prepared, providing a reliable foundation for subsequent model training.
[0146] Preferably, when the domain dataset is used to train the pre-trained language model, part of the parameter layers of the pre-trained language model are frozen. During training, the domain dataset only updates the parameters of another part of the parameter layers in the pre-trained language model. Then, when the updated pre-trained language model is trained using the annotated domain dataset, the model data is updated using the back propagation algorithm to obtain the supervised fine-tuning model.
[0147] When the reinforcement learning model is adjusted and reinforced, a KL divergence constraint with the supervised fine-tuning model is added to the loss function of the reinforcement learning model to prevent the reinforcement learning model from overfitting.
[0148] During the training of large language models, each graphics card is primarily occupied by two types of data: model state data and data other than model state. Model state data consists of model parameters (fp16), model gradients (fp16), and Adam state (fp32 momentum and fp32 variance). The remaining data consists of activation values, temporary buffers, and some graphics memory fragments. Among the model state parameters, optimizer state parameters account for the largest proportion. To optimize model state storage, ZeRO (Zero Redundancy Optimizer) technology is used to split the model parameters into multiple shards, allowing each worker node to store only a portion of the parameters. Specifically, the optimizer state is split 1 / N and distributed across N graphics cards. During forward and backward propagation, only the gradients are needed, not the optimizer. The remaining gradients are concatenated to form the complete gradient for backward propagation. After forward and backward propagation are complete, the optimizer is required to update the parameters. At this point, the optimizer parameters and complete gradients from multiple graphics cards are aggregated to perform the parameter update.
[0149] Figure 5 compares GPU memory usage with traditional training methods and after using ZeRO memory optimization technology. Generally speaking, in traditional data parallel processing, each process uses the same parameters for training and maintains a complete copy of the optimizer states, which also consumes a large amount of memory (1 Float16 = 2 bytes, 1 Float32 = 4 bytes). In a mixed-precision scenario, taking a model with Ψ as an example and the Adam optimizer, Adam needs to store: 1) a copy of the parameters and gradients in Float16 precision, each of which takes up 2Ψ bytes of memory; and 2) a copy of the parameters, momentum, and variance in Float32 precision, corresponding to three copies of 4Ψ, for a total memory usage of 2Ψ + 2Ψ + 3*4Ψ = 16Ψ bytes. Using ZeRO-1 technology, this can be converted to (2Ψ + 2Ψ + 3*4Ψ / n) bytes of memory. For example, when n = 3, only 8Ψ bytes of memory are required.
[0150] In addition, in order to improve training efficiency and solve the problem of floating-point underflow, mixed precision is used for training. During training, the gradients, parameters, and data used in forward propagation and backpropagation are all float16 type, and the optimizer saves a copy of float32 type parameters. When updating the parameters, the update results are saved as float32 type parameters, and then the float32 type parameters are converted to float16 type for forward propagation and backpropagation.
[0151] The formatting of the output instructions into control instructions is achieved through the output parsing module in the analysis management unit. The output parsing module receives the output instructions and parses them to form the feedback instructions and / or the control instructions. Feedback instructions are also generated in the output parsing module to interact with the user, facilitating further confirmation of the user's instruction target by the system. During this confirmation process, the user's operation guidance is completed until the correct control instruction is finally output. Of course, control instructions may also be issued occasionally during this process as needed. Please be aware that the monitoring, management, and control of photovoltaic power plants are generally not completed through a single control.
[0152] In one embodiment, the analysis and management unit further includes a memory storage module for storing photovoltaic standard data and real-time interaction data, and is connected to the context integration processing module. Specifically, the memory storage module includes a permanent storage submodule for storing photovoltaic standard data and a flash memory submodule for storing real-time interaction data. The photovoltaic standard data may include API call documentation, product specifications, photovoltaic industry technical information, etc., while the real-time interaction data may include the valid feature data and response status data corresponding to the execution of the control instructions.
[0153] The context comprehensive processing module extracts the data stored in the memory storage module based on the effective feature data to form a first input instruction. The first input instruction is a prompt word prompt. The first input instruction is fed back to the input preprocessing module, and the effective feature data and the first input instruction are aggregated in the input preprocessing module to form a comprehensive input instruction. The effective feature data and the prompt word prompt obtained after extracting the instruction input by the user are passed into the large language model. The prompt word prompt contains a list of control instructions and the output format of each control instruction (for example: access the power station list page {goPath: / station-list}, set the dark mode {setTheme:dark}, set the backup power to 30% {setBackupCapacity:30%}). The large language model has strong understanding and reasoning capabilities, and can complete the matching of the effective feature data and the prompt word prompt (mainly control instructions), and output the comprehensive input instruction in the agreed format.
[0154] The system also includes a user information management unit, which is connected to the context integration processing module. The user information management unit stores the user's account information and the device information bound to the account.
[0155] For example, after the instruction "I want to open a website" is input, as shown in Figure 6, the input preprocessing module extracts the user's account information and the valid voice feature data of "open a website", and the context comprehensive processing module calls the account information and the device information bound to it, and calls the product manual of the corresponding model device from the permanent storage sub-module to query and retrieve the relevant information about opening a website, and sends the account information and the device information bound to it and the website opening information as the first input instruction to the input preprocessing module. The input preprocessing module then sends the valid voice feature data of "open a website" together with the first input instruction as a comprehensive input instruction to the large language model, and then the large language model forms an output instruction based on natural language, and the output instructions are executed one by one (interaction of feedback instructions, issuance and execution of control instructions), and finally the website opening is completed.
[0156] In one embodiment, the system also includes a routing unit for distributing control instructions, which is configured in the cloud and connected to the analysis and management unit, and is used to distribute the control instructions to the terminal, cloud or device end for corresponding execution actions.
[0157] If the output parsing module receives the output instruction and parses the output instruction to form a control instruction, the system transmits the control instruction to the routing unit. The routing unit stores the configuration information of the call execution unit corresponding to each control instruction, and transmits the control instruction to the corresponding call execution unit according to the configuration information to execute the instruction (for example: goPath and setTheme instructions need to be transmitted to the interactive app to execute the corresponding jump or theme setting, and setBackupCapacity execution needs to be transmitted to the energy storage device on the device side to execute parameter setting).
[0158] In one embodiment, the system further includes a plurality of call execution units, each of which is configured at the terminal, cloud, or device end. Each call execution unit is connected to the routing unit, and the routing unit distributes control instructions to the call execution unit at the terminal, cloud, or device end to perform the corresponding execution action. The call execution unit is an application programming interface (API) connected to the terminal, cloud, or device end to implement operations on the terminal, cloud, or device end.
[0159] In one embodiment, the system further includes an action response unit, which is configured at the terminal. The action response unit is used to monitor the execution status of the action executed at the terminal, cloud or device end. Specifically, the input end of the action response unit is connected to each call execution unit for performing the execution action, and the output end of the action response unit is connected to the input preprocessing module. The call execution unit determines to call the terminal, cloud or device end to perform the corresponding execution action according to the control instruction. In the process of executing the action, the action response unit feeds back the response status data containing the execution status to the context comprehensive processing module, and recursively performs the execution action as needed.
[0160] In addition, in order to better manage the response status data, a storage management submodule is also provided in the memory storage module. The storage management submodule is connected to the action response unit. The storage management submodule receives the response status data fed back by the action response unit and updates it to the real-time interaction data. The real-time interaction data includes valid feature data and response status data.
[0161] It should be understood that large language models usually have several different role positionings, such as installers, after-sales technicians, electricity management personnel, customer service, intelligent assistants, intelligent customer service, etc. For example, when opening a site, the large language model plays the role of an "installer". In order to be able to realize the role positioning of the large language model when the user inputs instructions, it also includes an identity setting unit, which is connected to the context comprehensive processing module. The identity setting unit sets the working scene and role of the large language model according to external input or the valid feature data. External input is what is usually called manual setting, and the setting after the valid feature data is captured is automatic setting. For example, if the feature of "opening a site" (also known as the wake-up word) is captured, the large language model is automatically set to an installer. In addition, a fine-tuning module can be set to adjust the large language model so that the large language model can be more suitable for the current scenario.
[0162] It is also noted that the connection relationships described above include but are not limited to electrical connections and communication connections, and as long as they can complete the transmission of data and signals, they should all be protected by this application.
[0163] Example 2:
[0164] This embodiment provides a method for interactive management of a photovoltaic power station. This method is based on the system framework described in Example 1 and includes the following steps:
[0165] The user inputs a command through the interactive APP at the terminal. In this embodiment, the command is a voice command, and the input command is sent to the analysis and management unit at the cloud;
[0166] The instructions are analyzed and processed by the analysis management unit to form output instructions based on natural language, and the output instructions are formatted and parsed to form feedback instructions or executable control instructions;
[0167] The feedback instruction is sent back to the interactive APP to form an interactive communication with the user. During the interactive communication process, the user's new input instruction is combined with the feedback instruction to form new effective feature data, and then the data is searched to form a new comprehensive input instruction and sent to the large language model for processing;
[0168] The executable control instructions are in a formatted and parsed machine language format and are distributed to a terminal, a cloud, or a device installed at a photovoltaic power station according to the specific content of the control instructions to perform corresponding execution actions.
[0169] In one embodiment, the process of analyzing and processing the instruction at the analysis management unit to form an output instruction based on natural language includes:
[0170] Pre-process the instruction and capture the key words in the instruction to form effective feature data. Taking the website opening as an example, the key word is undoubtedly "website opening";
[0171] Searching for data from a memory storage module according to the valid feature data to form a first input instruction, and combining the first input instruction with the valid feature data to form a comprehensive input instruction;
[0172] The comprehensive input instructions are read and processed by a large language model to form output instructions based on natural language.
[0173] The valid feature data and prompt word obtained after extracting the instructions entered by the user are passed into the large language model. The prompt word prompt contains a list of control instructions and the output format of each control instruction (for example: access the power station list page {goPath: / station-list}, set dark mode {setTheme:dark}, set the backup power to 30% {setBackupCapacity:30%}). The large language model has strong understanding and reasoning capabilities, and can complete the matching of the valid feature data and the prompt word prompt (mainly control instructions), and output the comprehensive input instructions in the agreed format.
[0174] After the command "I want to open a website" is entered, the user's account information and the valid voice feature data of "open a website" are first extracted, and then the account information and the device information bound to it are called. According to the device information, the product manual of the corresponding model device is called to query and retrieve the relevant information for opening a website. The account information and the device information bound to it and the website opening information are used as the first input command, and then the valid voice feature data of "open a website" and the first input command are sent to the large language model as a comprehensive input command. The large language model then forms output commands based on natural language, and the output commands are executed one by one (interaction of feedback commands, issuance and execution of control commands), and finally the website opening is completed.
[0175] In one embodiment, the process of searching for data from the memory storage module according to the valid feature data to form the first input instruction includes:
[0176] Based on the valid characteristic data, data is searched from a memory storage module to search for photovoltaic standard data and / or real-time interactive data corresponding to the valid characteristic data, and the searched data is aggregated to form the first input instruction. The search for photovoltaic standard data and / or real-time interactive data corresponding to the valid characteristic data utilizes a support vector search method.
[0177] It should also be noted that, for the executable control instructions, when distributed to the terminal, cloud, or equipment installed at the photovoltaic power station based on the specific content of the control instructions, the corresponding execution action may include multiple subtasks. The execution action with multiple subtasks is disassembled, and the system recursively executes each subtask automatically.
[0178] There are two ways to set the working scenarios and roles of large language models: automatic setting and manual setting.
[0179] During the automatic setting process, after the instruction forms valid feature data through analysis and processing at the analysis management unit, the working scene and role of the large language model are automatically set according to the valid feature data.
[0180] During the manual setting process, the user manually inputs identity setting parameters when inputting instructions through the interactive APP at the terminal, and manually sets the working scenario and role of the large language model.
[0181] In addition, when the analysis management unit searches for data from the memory storage module based on the valid feature data, it simultaneously retrieves the user's account information and the device information bound to the account, and summarizes the account information and device information together in the first input instruction.
[0182] It can be seen that the innovative introduction of a large language model (such as the ChatGPT model) uses comprehensive input instructions in text form on the input side, and output instructions based on natural language on the output side. From front-end input to back-end processing and output, all are based on natural language processing, which is more in line with the usage needs of traditional users. During the final specific execution, the natural language-based output instructions are converted into a format for specific execution, which reduces the processing difficulty of this management method. In addition, the data sources in the processing process are more diverse, making the system processing capability stronger and the method more comprehensive in its ability to handle problems.
[0183] Based on the detailed description of the specific embodiments of the present application in conjunction with the accompanying drawings below, those skilled in the art will become more aware of the above and other objects, advantages and features of the present application.
[0184] An embodiment of the present application also provides a computing device. Referring to Figure 2, the computing device includes a memory 1120, a processor 1110, and a computer program stored in the memory 1120 and capable of being run by the processor 1110. The computer program is stored in a space 1130 for program code in the memory 1120. When the computer program is executed by the processor 1110, it is implemented to execute any one of the method steps 1131 according to the present invention.
[0185] The embodiment of the present application further provides a computer-readable storage medium. Referring to FIG3 , the computer-readable storage medium includes a storage unit for program code, wherein the storage unit is provided with a program 1131 ′ for executing the method steps according to the present invention, and the program is executed by a processor.
[0186] The present application also provides a computer program product comprising instructions. When the computer program product is run on a computer, the computer is caused to execute the steps of the method according to the present invention.
[0187] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When software is used for implementation, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer loads and executes the computer program instructions, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrations. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).
[0188] Professionals should also be further aware that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0189] Those skilled in the art will appreciate that all or part of the steps in the above-described embodiment method can be performed by instructing a processor through a program, and the program can be stored in a computer-readable storage medium, which is a non-transitory medium, such as a random access memory, a read-only memory, a flash memory, a hard disk, a solid-state drive, a magnetic tape, a floppy disk, an optical disc, and any combination thereof.
[0190] In particular, the present invention also provides a computing device, comprising a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor implements the method described above when executing the computer program.
[0191] The present invention also provides a computer-readable storage medium having a computer program stored therein, wherein the computer program implements the above method when executed by a processor. The computer-readable storage medium is preferably a non-volatile readable storage medium.
[0192] The above embodiments are only for illustrating the technical concept and features of the present invention. Their purpose is to enable people familiar with this technology to understand the contents of the present invention and implement them accordingly. They are not intended to limit the scope of protection of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the scope of protection of the present invention.
Claims
1. A photovoltaic power station interactive management system, characterized in that: include: A terminal is equipped with an interactive APP, and the terminal receives instructions input by a user; The equipment end is installed at the photovoltaic power station, and the equipment end controls and manages the photovoltaic power station; The cloud is connected to the terminal and the device in communication; An analysis management unit, which is configured at the terminal, device or cloud, receives the instruction and performs analysis and processing to form an output instruction, which forms a feedback instruction or an executable control instruction after being parsed; Among them, the feedback instruction is sent back to the interactive APP to form interactive communication with the user, and the control instruction is distributed to the terminal, cloud or device to perform corresponding execution actions.
2. The photovoltaic power station interactive management system according to claim 1, characterized in that: The analysis management unit comprises: An input preprocessing module is used to receive the instruction and perform feature capture on the instruction to form valid feature data in text form; A context comprehensive processing module is connected to the input preprocessing module, receives the effective feature data, and retrieves and calls the prompt word prompt matching the effective feature data to form a comprehensive input instruction; The large language model has an input side connected to the input preprocessing module and an output side connected to the output parsing module, receives the comprehensive input instructions and processes them to form output instructions, and the output instructions are output instructions based on natural language.
3. The photovoltaic power station interactive management system according to claim 2, characterized in that: The analysis management unit also includes a memory storage module, which is used to store photovoltaic standard data and real-time interactive data. The memory storage module is connected to the context comprehensive processing module.
4. The photovoltaic power station interactive management system according to claim 3, characterized in that: The context comprehensive processing module extracts the data stored in the memory storage module according to the valid feature data to form a first input instruction, wherein the first input instruction is a prompt word, and feeds the first input instruction back to the input preprocessing module, wherein the valid feature data and the first input instruction are aggregated in the input preprocessing module to form a comprehensive input instruction.
5. The photovoltaic power station interactive management system according to claim 3, characterized in that: The memory storage module includes a permanent storage submodule and a flash memory submodule. The permanent storage submodule is used to store photovoltaic standard data, and the flash memory submodule is used to store real-time interactive data.
6. The photovoltaic power station interactive management system according to claim 2 or 5, characterized in that: The photovoltaic standard data includes API call documents, product descriptions, and photovoltaic industry technical information.
7. The photovoltaic power station interactive management system according to claim 2, characterized in that: The analysis management unit also includes an output parsing module, the output side of the large language model is connected to the output parsing module, and the output parsing module receives the output instruction and parses the output instruction to form the feedback instruction and / or the control instruction.
8. The photovoltaic power station interactive management system according to claim 2, characterized in that: The large language model includes a ChatGPT model.
9. The photovoltaic power station interactive management system according to claim 2, characterized in that: The establishment of the large language model includes the following training process: Acquire photovoltaic-related basic data and instruction data to construct a domain data set, and annotate the domain data set to obtain an annotated domain data set; The pre-trained language model is trained by the domain dataset to update the pre-trained language model, and the updated pre-trained language model is trained by the labeled domain dataset to obtain a supervised fine-tuning model; The supervised fine-tuning model adds a classification mechanism, and then combines manual feedback to the output of the supervised fine-tuning model after the classification is added for scoring and ranking, thereby forming a reward model with a scoring mechanism; The domain dataset is used as input to train the reinforcement learning model, and the output is sent to The reward model is used to score, and the obtained answer score is fed back to the reinforcement learning model for adjusting and strengthening the reinforcement learning model; The obtained adjusted and reinforced reinforcement learning model is the final large language model.
10. The photovoltaic power station interactive management system according to claim 9, characterized in that: When the domain dataset is used to train the pre-trained language model, part of the parameter layers of the pre-trained language model are frozen, and when training is performed, the domain dataset only updates parameters of another part of the parameter layers in the pre-trained language model.
11. The photovoltaic power station interactive management system according to claim 9 or 10, characterized in that: When the annotated domain dataset is used to train the updated pre-trained language model, the model data is updated through a back-propagation algorithm to obtain the supervised fine-tuning model.
12. The photovoltaic power station interactive management system according to claim 9, characterized in that: When the reinforcement learning model is adjusted and reinforced, a KL divergence constraint with the supervised fine-tuning model is added to the loss function of the reinforcement learning model to prevent the reinforcement learning model from overfitting.
13. The photovoltaic power station interactive management system according to claim 9, characterized in that: The photovoltaic-related basic data includes one or more of the user manual of the photovoltaic power station, frequently asked questions and answers, technical documents, network materials, and user feedback information. The instruction data is obtained by parsing the system instruction set and interface parameters of the photovoltaic power station.
14. The photovoltaic power station interactive management system according to claim 1, characterized in that: The system also includes a routing unit, which is connected to the analysis and management unit and is used to distribute the control instructions to the terminal, cloud or device end to perform corresponding execution actions.
15. The photovoltaic power station interactive management system according to claim 14, characterized in that: The system also includes multiple call execution units, which are respectively configured at the terminal, cloud or device end, each call execution unit is connected to the routing unit, and the routing unit distributes control instructions to the call execution unit at the terminal, cloud or device end to perform corresponding execution actions.
16. The photovoltaic power station interactive management system according to claim 2, characterized in that: The system also includes an action response unit, which is used to monitor the execution status of the action executed at the terminal, cloud or device.
17. The photovoltaic power station interactive management system according to claim 16, characterized in that: The input end of the action response unit is connected to each calling execution unit used to perform the execution action, and the output end of the action response unit is connected to the input preprocessing module. The calling execution unit is respectively configured at the terminal, cloud or device end, and the calling execution unit determines to call the terminal, cloud or device end to perform the corresponding execution action according to the control instruction. During the execution of the action, the action response unit feeds back the response status data containing the execution status to the context comprehensive processing module, and recursively performs the execution action as needed.
18. The photovoltaic power station interactive management system according to claim 16, characterized in that: The memory storage module includes a storage management submodule, which is connected to the action response unit. The storage management submodule receives the response status data fed back by the action response unit and updates it into the real-time interaction data.
19. The photovoltaic power station interactive management system according to claim 2, characterized in that: It also includes an identity setting unit, which is connected to the context comprehensive processing module. The identity setting unit sets the working scene and role of the large language model according to external input or the effective feature data.
20. The photovoltaic power station interactive management system according to claim 2, characterized in that: It also includes a user information management unit, which is connected to the context comprehensive processing module. The user information management unit stores the user's account information and the device information bound to the account.
21. The photovoltaic power station interactive management system according to claim 1, characterized in that: The instruction input by the user is voice or text input.
22. A photovoltaic power station interactive management method, characterized in that: The steps include: The user inputs instructions through the interactive APP at the terminal, and sends the input instructions to the analysis management unit; The instruction is analyzed and processed at the analysis management unit to form an output instruction, and then the output instruction is formatted and parsed to form a feedback instruction or an executable control instruction. The analysis management unit is configured at the terminal, device or cloud; For the feedback instruction, sending it back to the interactive APP to form interactive communication with the user; For the executable control instruction, the control instruction is in a machine language format after formatting and parsing, and is distributed to the terminal, the cloud or the equipment end installed in the photovoltaic power station according to the specific content of the control instruction to perform the corresponding execution action.
23. The photovoltaic power station interactive management method according to claim 22, characterized in that: The process of forming an output instruction by analyzing and processing the instruction at the analysis management unit includes: Preprocessing the instruction to form effective feature data in text form by capturing features of keywords in the instruction; Searching for data from a memory storage module according to the valid feature data to form a first input instruction, wherein the first input instruction is a prompt word prompt, and combining the first input instruction with the valid feature data to form a comprehensive input instruction; The comprehensive input instructions are read and processed by a large language model to form output instructions.
24. The photovoltaic power station interactive management method according to claim 23, characterized in that: The process of searching for data from a memory storage module according to valid feature data to form a first input instruction includes: According to the effective characteristic data, data is searched from the memory storage module, photovoltaic standard data and / or real-time interactive data corresponding to the effective characteristic data are searched, and the searched data are summarized to form the First input instruction.
25. The photovoltaic power station interactive management method according to claim 24, characterized in that: The process of searching for photovoltaic standard data and / or real-time interactive data corresponding to the effective characteristic data adopts a support vector search method.
26. The photovoltaic power station interactive management method according to claim 22, characterized in that: The feedback instruction is sent back to the interactive communication process between the interactive APP and the user. The new input instructions of the user are combined with the feedback instructions to form new effective feature data, and then the data search process is performed to form a new comprehensive input instruction and send it to the large language model for processing.
27. The photovoltaic power station interactive management method according to claim 22, characterized in that: The executable control instructions are distributed to the terminal, the cloud or the equipment installed at the photovoltaic power station according to the specific content of the control instructions, and the corresponding execution actions are performed. For an execution action with multiple subtasks, the execution action is disassembled, and the system recursively executes each subtask automatically.
28. The photovoltaic power station interactive management method according to claim 22, characterized in that: After the analysis management unit forms valid feature data through analysis and processing, the instruction automatically sets the working scene and role of the large language model according to the valid feature data.
29. The photovoltaic power station interactive management method according to claim 22, characterized in that: When the user inputs instructions through the interactive APP at the terminal, the user manually inputs identity setting parameters to manually set the working scene and role of the large language model.
30. The photovoltaic power station interactive management method according to claim 22, characterized in that: When the analysis management unit searches for data from the memory storage module according to the valid feature data, it simultaneously retrieves the user's account information and the device information bound to the account, and summarizes the account information and the device information together in the first input instruction.
31. An analysis management unit, characterized in that: The analysis management unit forms an interactive communication with the user through the interactive APP. The analysis management unit receives the instructions input by the user and performs analysis and processing to form an output instruction. The output instruction forms a feedback instruction or an executable control instruction after parsing. The feedback instruction is sent back to the interactive APP to form an interactive communication with the user, and the control instruction is distributed to the terminal, cloud or device end to perform the corresponding execution action; The analysis management unit comprises: An input preprocessing module is used to receive the instruction and capture features of the instruction to form effective feature data; A context comprehensive processing module, connected to the input preprocessing module, receives the valid feature data, and retrieves and calls data matching the valid feature data to form a comprehensive input instruction; The input side of the large language model is connected to the input preprocessing module, receives the comprehensive input instructions and processes them to form output instructions.
32. The analysis management unit according to claim 31, characterized in that: The analysis management unit also includes a memory storage module, which is used to store photovoltaic standard data and real-time interactive data. The memory storage module is connected to the context comprehensive processing module, wherein the photovoltaic standard data includes API call documents, product descriptions, and photovoltaic industry technical information.
33. The analysis management unit according to claim 32, characterized in that: The memory storage module includes a permanent storage submodule and a flash memory submodule. The permanent storage submodule is used to store photovoltaic standard data, and the flash memory submodule is used to store real-time interactive data.
34. The analysis management unit according to claim 32 or 33, characterized in that: The context comprehensive processing module extracts the data stored in the memory storage module according to the effective feature data to form a first input instruction, and feeds the first input instruction back to the input preprocessing module, where the effective feature data and the first input instruction are aggregated to form a comprehensive input instruction.
35. The analysis management unit according to claim 31, characterized in that: The analysis management unit also includes an output parsing module, the output side of the large language model is connected to the output parsing module, and the output parsing module receives the output instruction and parses the output instruction to form the feedback instruction and / or the control instruction.
36. The analysis management unit according to claim 31, characterized in that: The establishment of the large language model includes the following training process: Acquire photovoltaic-related basic data and instruction data to construct a domain data set, and annotate the domain data set to obtain an annotated domain data set; The pre-trained language model is trained by the domain dataset to update the pre-trained language model, and the updated pre-trained language model is trained by the labeled domain dataset to obtain a supervised fine-tuning model; The supervised fine-tuning model adds a classification mechanism, and then combines manual feedback to the output of the supervised fine-tuning model after the classification is added for scoring and ranking, thereby forming a reward model with a scoring mechanism; The domain dataset is used as input to train the reinforcement learning model, the obtained output result is sent to the reward model for scoring, and the obtained answer score is fed back to the reinforcement learning model for adjusting and strengthening the reinforcement learning model; The obtained adjusted and reinforced reinforcement learning model is the final large language model.
37. The analysis management unit according to claim 36, characterized in that: When the domain dataset is used to train the pre-trained language model, part of the parameter layers of the pre-trained language model are frozen, and when training is performed, the domain dataset only updates parameters of another part of the parameter layers in the pre-trained language model.
38. The analysis management unit according to claim 36 or 37, characterized in that: When the annotated domain dataset is used to train the updated pre-trained language model, the model data is updated through a back-propagation algorithm to obtain the supervised fine-tuning model.
39. The analysis management unit according to claim 36, characterized in that: When the reinforcement learning model is adjusted and reinforced, a KL divergence constraint with the supervised fine-tuning model is added to the loss function of the reinforcement learning model to prevent the reinforcement learning model from overfitting.
40. The analysis management unit according to claim 36, characterized in that: The photovoltaic-related basic data includes one or more of the user manual of the photovoltaic power station, frequently asked questions and answers, technical documents, network materials, and user feedback information. The instruction data is obtained by parsing the system instruction set and interface parameters of the photovoltaic power station.
Citation Information
Patent Citations
Vehicle speech interaction method and system and computer readable storage medium
CN107204185A
Photovoltaic power station operation and maintenance control method, device and system
CN114328964A
Intelligent scheduling maintenance identification library construction method
CN114996470A
Natural language instruction information processing method and device, equipment and medium
CN115756610A
Using large language model(s) in generating automated assistant response(s
US20230074406A1
Cited By
Analysis management unit and photovoltaic power station interactive management system
CN118885561A
Power grid analysis tool calling method and system based on large language model
CN119336796A
Human-machine natural interaction method and system based on body intelligence
CN120669865A
Processing method and system for presenting double intelligent agents based on service disassembly and visualization
CN120744108A
Large model guided multi-energy microgrid risk scheduling method and system
CN122371352A
