Energy consumption prediction and management method based on large model agent

Abstract

The invention belongs to the technical field of building energy consumption prediction, and discloses an energy consumption prediction and management method based on a large model agent, which comprises the following steps of: obtaining a building energy consumption data set and constructing memory of the agent; the intelligent agent interacts with the environment in real time and updates memory; the intelligent agent carries out planning, and a large-scale task is decomposed into sub-tasks; the intelligent agent uses a tool to call external resources; the intelligent agent acts to predict and manage the energy consumption of the building; and updating the environment in real time according to feedback generated by the action. The building energy consumption prediction and management system is constructed on the basis of a large model agent technology, various data types such as texts, images and time sequences can be processed at the same time, powerful time sequence modeling and global dependency capture capabilities are achieved, manual intervention is not needed, important features in data are automatically learned, and the system can adapt to different buildings and conditions and is high in practicability. Online learning and continuous learning can be realized, energy consumption management is performed while prediction is performed, and the energy consumption is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of building energy consumption prediction, and in particular relates to an energy consumption prediction and management method based on a large model intelligent body. Background Art

[0002] In recent years, with the development of artificial intelligence, Internet of Things and other related technologies, smart cities have attracted more and more attention from local governments and academia. As an important part of smart city construction, intelligent buildings refer to a building application model that uses communication technology as the main trunk and uses system integration methods to organically combine computer technology, network technology, automatic control technology, software engineering technology and architectural art design, break through the information barriers between isolated systems, make buildings an intelligent subject with information exchange, and realize intelligent management of buildings and effective use of their information resources.

[0003] Artificial intelligence related algorithms, such as neural networks, Bayesian estimation, K-means, Markov, expert systems, etc., are widely used in related fields of intelligent buildings. A typical example is the application of knowledge-based expert systems for the detection of building air conditioning water system faults. Through modeling, expert knowledge is input, training samples are trained, and finally a practical real-time fault detection system is generated through training and learning. However, traditional artificial intelligence algorithms have some unavoidable problems in the energy consumption prediction and management of intelligent building groups. For example, building energy consumption prediction usually requires processing multimodal data (such as time series, sensor data, meteorological data, text, images, etc.). Traditional artificial intelligence methods are usually only good at processing a single type of data. When processing multimodal data, complex data fusion and preprocessing processes are required, which are prone to errors and inefficient. For example, Chinese patent CN118709721A discloses a method for building energy consumption prediction using a combined model, but can only process single-modal data information. Building energy consumption changes dynamically with time, environment and user behavior. It is difficult to cope with dynamic changes. Traditional artificial intelligence models need to be frequently retrained or adjusted when facing dynamic changes, and cannot adapt to rapidly changing environments. For example, Chinese patent CN118569680A discloses a method for predicting building energy consumption using a clustering method, but it is difficult to update online. Summary of the invention

[0004] In view of the above problems, the purpose of the present invention is to provide a building energy consumption prediction and management method based on a large model intelligent agent, which can efficiently process multimodal data, has strong generalization ability, long-term adaptation and continuous learning.

[0005] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0006] A method for energy consumption prediction and management based on a large model intelligent agent comprises the following steps:

[0007] Step 1: Obtain the building energy consumption dataset and build the agent’s memory;

[0008] Step 2: The agent interacts with the environment in real time and updates its memory;

[0009] Step 3: The agent performs planning and breaks down large tasks into subtasks;

[0010] Step 4: The agent uses tools to call external resources;

[0011] Step 5: The intelligent agent takes action to predict and manage building energy consumption;

[0012] Step 6: Update the environment in real time based on the feedback generated by the action.

[0013] Furthermore, the data described in Step 1 includes one or more of building number, time, electric meter reading, building area, air temperature, wind speed, wind direction, sea level pressure, and satellite cloud map.

[0014] Furthermore, the data needs to be preprocessed to fill in missing values.

[0015] Furthermore, the environmental interaction described in Step 2 includes obtaining real-time data through sensors in the building as the basis for the agent's decision-making; integrating external information including weather forecasts, energy prices, and user behavior patterns into the agent to more accurately predict energy consumption needs.

[0016] Furthermore, the sensor includes one or more of a temperature sensor, a humidity sensor, an energy consumption meter, and a light sensor.

[0017] Furthermore, in Step 3, the agent conducts self-criticism and self-reflection on its past behaviors, learns from its mistakes and makes improvements for future steps; ReAct technology is used to integrate reasoning and action into the agent by expanding the action space into a combination of task-specific discrete actions and language space.

[0018] Furthermore, the ReAct technique is used to integrate reasoning and action into the agent by expanding the action space into a combination of task-specific discrete actions and language space, including the following steps:

[0019] Step 3-1: Expand the action space of the agent to Where L is the language space and A is the original action space.

[0020] Step3-2: is an action in the language space, called a thought, which uses the current context information c t Inference, where c t =(o 1 , a 1 , ..., o t-1 , a t-1 , o t ), a t ∈A is the action at time t, o t is the observed value from the environment at time t;

[0021] Step 3-3: Update context information, To support subsequent reasoning.

[0022] Furthermore, in Step 4, the agent calls the external API to obtain additional information that is missing in the large model, including code execution capabilities and access to proprietary information sources.

[0023] Furthermore, the agent calls external APIs to obtain additional information that is missing from the large model, including code execution capabilities and access to proprietary information sources. The specific operation process is as follows:

[0024] Step 4-1 Task planning: parse the total task into multiple tasks;

[0025] Step 4-2 Model selection, the agent assigns the task to multiple expert models;

[0026] Step 4-3 Task execution: the expert model performs specific tasks and records the results;

[0027] Step 4-4: Response generation, the agent receives and summarizes the execution results.

[0028] Furthermore, in Step 6, the intelligent agent updates its model and control strategy by continuously collecting feedback information including prediction errors and energy consumption changes to adapt to the changing needs of the building; as time goes by, the intelligent agent will continuously optimize its control strategy through a large amount of data and even adapt to changes in the building's usage functions.

[0029] The present invention has the following beneficial effects:

[0030] Based on large-model intelligent agent technology, a building energy consumption prediction and management system is constructed. It can process multiple data types such as text, images, time series, etc. at the same time. It has powerful time series modeling and global dependency capture capabilities. It does not require human intervention and can automatically learn important features in the data. It has strong generalization capabilities, can adapt to different buildings and conditions, can learn online and continuously, and manage energy consumption while predicting, thereby reducing energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of the building energy consumption prediction and management method based on large model intelligent body in the present invention.

[0032] Figure 2 This is the overall architecture diagram of the large model intelligent body in the present invention.

[0033] Figure 3 It is a memory structure diagram in the present invention. DETAILED DESCRIPTION

[0034] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0035] The present invention constructs a building energy consumption prediction and management system based on a large model intelligent agent. After the system is deployed, it interacts with the environment in real time, can learn and adapt to changing data online, and utilizes the capabilities of the large model to perform task planning and use external tools, etc., to improve the prediction and management capabilities.

[0036] like Figure 1-3 As shown, the building energy consumption prediction and management method based on the large model intelligent agent includes the following steps:

[0037] Step 1: Obtain the building energy consumption dataset and construct the memory of the intelligent agent. The multimodal data includes building number, time, meter reading, building area, air temperature, wind speed, wind direction, sea level pressure, satellite cloud map, etc., and perform preprocessing such as missing value filling.

[0038] Step 2: The intelligent agent interacts with the environment in real time and updates its memory. Environmental interaction includes obtaining real-time data through sensors in the building (such as temperature, humidity, energy consumption meters, light sensors, etc.) as the basis for the intelligent agent's decision-making; integrating external weather forecasts, energy prices, user behavior patterns and other information into the intelligent agent to more accurately predict energy consumption needs.

[0039] Among them, memory updating includes the following steps:

[0040] (1) Memory writing:

[0041] Given a task k, if the agent takes action at step t And the environment provides observation Then the memory write operation can be expressed as:

[0042]

[0043] Where W is a projection function, It is the memory content that is finally written and stored.

[0044] (2) Memory management:

[0045] make is the memory content of task k before step t, is the information written in step t, then the memory management operation is:

[0046]

[0047] Where P is a function that iteratively processes memory information.

[0048] (3) Memory reading:

[0049] Assumptions is the memory content of task k at step t, is the context of the next action, then the memory read operation can be expressed as:

[0050]

[0051] Where R is and The similarity between As part of the final prompt, it drives the big model to take the next action.

[0052] Based on the above operations, we can get a arrive The comprehensive formula is:

[0053]

[0054] Among them, LLM is the large model. The above formula represents the interaction process between the large model agent and the environment through iteration.

[0055] Step 3: The agent performs planning and breaks down large tasks into subtasks. The agent can self-criticize and self-reflect on past behaviors, learn from mistakes and make improvements for future steps, thereby improving the quality of the final result. ReAct technology is used to integrate reasoning and action into the agent by expanding the action space into a combination of task-specific discrete actions and language space, including the following steps:

[0056] Step 3-1: Expand the action space of the agent to Where L is the language space and A is the original action space.

[0057] Step3-2: is an action in the language space, called a thought, which uses the current context information c t Inference, where c t =(o 1 , a 1 , ..., ot-1 , a t-1 , o t ), a t ∈A is the action at time t, o t is the observation from the environment at time t.

[0058] Step 3-3: Update context information, To support subsequent reasoning.

[0059] Step 4: The agent uses tools to call external resources. The agent calls external APIs, knowledge bases, or models to obtain additional information that is missing from the large model, including code execution capabilities, access to proprietary information sources, etc. The following steps are included:

[0060] Step 4-1 Task planning: parse the total task into multiple tasks.

[0061] Step 4-2 Model selection: the agent assigns tasks to multiple expert models.

[0062] Step 4-3 Task execution: the expert model performs specific tasks and records the results.

[0063] Step 4-4: Response generation, the agent receives and summarizes the execution results.

[0064] Step 5: The intelligent agent takes action to predict and manage building energy consumption;

[0065] Step 6: Update the environment in real time based on the feedback generated by the actions. The intelligent agent continuously collects feedback information such as prediction errors and energy consumption changes, updates its model and control strategy, and adapts to the changing needs of the building. Over time, the intelligent agent will continuously optimize its control strategy through a large amount of data, and even adapt to changes in the building's use function (such as seasonal changes, building expansion, etc.).

[0066] After each prediction, the reinforcement learning PPO method is used to fine-tune the large model based on the prediction results and the true value to achieve the update of the large model:

[0067] Let the objective function be:

[0068]

[0069] in,

[0070]

[0071] represents the probability ratio;

[0072] represents the estimated value of the advantage function at step t;

[0073] γ is the discount factor;

[0074] V(s) represents the state value function;

[0075] represents the squared difference loss;

[0076] S represents the entropy reward.

[0077] At each iteration, the parameters are updated to maximize the objective function.

[0078] This invention relies on advanced large-scale intelligent agent technology as the core driving force to innovatively build a comprehensive and efficient building energy consumption prediction and management system. This system not only deeply integrates the latest research results in the field of artificial intelligence, but also achieves significant technological progress and breakthroughs in data processing, model construction and practical application.

[0079] 1. In-depth analysis of technical theory

[0080] First, at the data processing level, this system demonstrates unprecedented inclusiveness and flexibility. It cleverly integrates three key technologies: natural language processing (NLP), computer vision (CV), and time series analysis, and realizes all-round, multi-dimensional processing of text records (such as building usage logs, maintenance records), image data (such as building appearance, internal layout photos), and continuous time series data (such as ambient temperature, electricity consumption, etc.). This cross-modal data integration capability lays a solid foundation for subsequent accurate prediction and scientific management.

[0081] 2. Innovation in Timing Modeling and Global Dependency Capture

[0082] In terms of model construction, this system uses advanced sequence models in deep learning, such as long short-term memory networks (LSTM) and Transformer, which are known for their excellent time series modeling capabilities and can capture complex long-term dependencies between data. More importantly, this system further enhances the ability to capture global dependencies by optimizing algorithm design, so that the model can cross time boundaries when understanding building energy consumption patterns, comprehensively consider historical, current and even possible environmental changes and usage habits in the future, and thus make more accurate predictions.

[0083] 3. A leap in automated learning and generalization capabilities

[0084] Different from the limitations of traditional methods that rely on a lot of manual intervention and preset rules, this system realizes the full automation from data collection, feature extraction to model training. With the help of advanced unsupervised learning and semi-supervised learning technology, the system can automatically learn the key features in the data without manual labeling, greatly reducing the cost and time of data preparation. At the same time, this automated learning mechanism gives the system a strong generalization ability. Whether it is facing buildings of different types and sizes, or dealing with complex and changeable energy consumption conditions such as seasonality and suddenness, the system can quickly adapt and respond accurately.

[0085] 4. Advantages of online learning and continuous optimization

[0086] What is more worth mentioning is that this system supports online learning and continuous learning mechanisms. This means that as new data continues to flow in, the system can update its internal model in real time and continuously optimize the prediction accuracy and management strategy. This dynamic adjustment capability enables the system to keep up with the actual trend of changes in building energy consumption and always maintain the effectiveness and cutting-edge nature of prediction and management. In addition, by integrating advanced reinforcement learning technology, the system can also autonomously explore the optimal strategy in the process of energy consumption management to achieve further reduction in energy consumption and efficient use of resources.

[0087] In summary, the building energy consumption prediction and management system constructed by the present invention has achieved comprehensive innovation in cross-modal data processing, advanced time series modeling, automated learning and continuous optimization in technical theory, significantly improving the intelligence level and practical application effect of the system. It not only provides strong technical support for energy conservation and emission reduction in the construction industry, but also opens up broad space for the construction and development of future smart cities.

[0088] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for energy consumption prediction and management based on a large model intelligent agent, characterized in that: The following steps are involved: Step 1: Obtain the building energy consumption dataset and build the agent’s memory; Step 2: The agent interacts with the environment in real time and updates its memory; Step 3: The agent performs planning and breaks down large tasks into subtasks; Step 4: The agent uses tools to call external resources; Step 5: The intelligent agent takes action to predict and manage building energy consumption; Step 6: Update the environment in real time based on the feedback generated by the action.

2. The energy consumption prediction and management method based on a large model agent according to claim 1 is characterized in that: The data described in Step 1 includes one or more of building number, time, electric meter reading, building area, air temperature, wind speed, wind direction, sea level pressure, and satellite cloud map.

3. The energy consumption prediction and management method based on large model intelligent agent according to claim 2 is characterized in that: The data needs to be preprocessed to fill in missing values.

4. The energy consumption prediction and management method based on a large model agent according to claim 1 is characterized in that: The environmental interaction described in Step 2 includes obtaining real-time data through sensors in the building as the basis for the agent's decision-making; integrating external information including weather forecasts, energy prices, and user behavior patterns into the agent to more accurately predict energy consumption needs.

5. The energy consumption prediction and management method based on large model intelligent agent according to claim 4 is characterized in that: The sensor includes one or more of a temperature sensor, a humidity sensor, an energy consumption meter, and a light sensor.

6. The energy consumption prediction and management method based on large model intelligent agent according to claim 1 is characterized in that: In Step 3, the agent conducts self-criticism and self-reflection on past behaviors, learns from mistakes and makes improvements for future steps; ReAct technology is used to integrate reasoning and action into the agent by expanding the action space into a combination of task-specific discrete actions and language space.

7. The energy consumption prediction and management method based on large model intelligent agent according to claim 6 is characterized in that: The ReAct technique is used to integrate reasoning and action into the agent by expanding the action space into a combination of task-specific discrete actions and language space, which includes the following steps: Step 3-1: Expand the action space of the agent to Where L is the language space and A is the original action space. Step3-2: is an action in the language space, called a thought, which uses the current context information c t Inference, where c t =(o1, a1, ..., o t-1 , a t-1 , o t ), a t ∈A is the action at time t, o t is the observed value from the environment at time t; Step 3-3: Update context information, To support subsequent reasoning.

8. The energy consumption prediction and management method based on large model intelligent agent according to claim 1 is characterized in that: In Step 4, the agent calls the external API to obtain additional information that is missing from the large model, including code execution capabilities and access to proprietary information sources.

9. The energy consumption prediction and management method based on large model intelligent agent according to claim 8 is characterized in that: The agent calls the external API to obtain additional information that is missing from the large model, including code execution capabilities and access to proprietary information sources. The specific operation process is as follows: Step 4-1 Task planning: parse the total task into multiple tasks; Step 4-2 Model selection, the agent assigns the task to multiple expert models; Step 4-3 Task execution: the expert model performs specific tasks and records the results; Step 4-4: Response generation, the agent receives and summarizes the execution results.

10. The energy consumption prediction and management method based on large model intelligent agent according to claim 1 is characterized in that: In Step 6, the intelligent agent continuously collects feedback information including prediction errors and energy consumption changes, updates its model and control strategy, and adapts to the changing needs of the building. As time goes by, the intelligent agent will continuously optimize its control strategy through large amounts of data and even adapt to changes in the building's usage functions.

Images