Building energy-saving self-adaptive regulation and control method based on scene analysis

Abstract

The invention provides a scene analysis-based building energy-saving self-adaptive regulation and control method. The method comprises the following steps of forming a state space for representing a building scene through time characteristic information, environment data and equipment operation state data; discretization or parameterization is carried out on executable regulation and control operation of building energy consumption equipment to be controlled to form an action space; the state space serves as input, the action space serves as output, the multi-target reward function serves as an optimization target, and a DQN model is trained; inputting the current state space into the pre-trained DQN model, and outputting a device control action by the DQN model; generating an equipment adjusting strategy according to the equipment control action, and executing the equipment adjusting strategy; obtaining a new state and actual energy consumption data of the building equipment after the equipment adjustment strategy is executed; the DQN model is adjusted through regular sampling, and the action space of the DQN model is updated. According to the method, the building operation scene can be dynamically constructed, the cross-equipment collaborative optimization control strategy is generated, and the energy consumption of the building is reduced.

Description

Technical Field

[0001] This invention relates to the field of building management technology, and in particular to a building energy-saving adaptive control method based on scenario analysis. Background Technology

[0002] Building energy-saving control technologies primarily rely on building automation systems. Traditional building control systems generally employ programmed control based on fixed schedules or feedback control strategies based on single parameter thresholds to independently start, stop, or regulate equipment such as air conditioning and lighting. In recent years, with the development of control theory, advanced methods such as proportional-integral-derivative control and fuzzy logic control have been introduced into this field to improve the adaptability of single control loops. Simultaneously, the widespread adoption of IoT technology has made it possible to collect real-time data on the building environment and equipment through multi-sensor networks, laying the foundation for data-driven monitoring and management.

[0003] However, existing technologies still face significant limitations and industry pain points in practical applications. Traditional control methods are inherently static or semi-static, making it difficult to dynamically respond to complex scenarios involving multiple coupled factors such as fluctuations in population density, weather changes, and equipment performance degradation, easily leading to energy waste or decreased comfort. More importantly, subsystems such as lighting, air conditioning, and ventilation often operate and are controlled independently, lacking collaborative optimization mechanisms, making it difficult to achieve globally optimal overall building energy efficiency. Although intelligent methods such as fuzzy control improve adaptability to some extent, they heavily rely on expert experience to build knowledge bases, resulting in inherent problems such as high design costs and difficulty in covering unknown scenarios. Furthermore, most existing systems are still open-loop or semi-closed-loop controls, with energy consumption analysis typically lagging behind, unable to perform online strategy optimization based on real-time operational feedback, and essentially lacking the adaptive capability for continuous self-learning and evolution. Summary of the Invention

[0004] This invention provides a building energy-saving adaptive control method based on scenario analysis. It can dynamically construct building operation scenarios through multimodal sensing data and generate cross-device collaborative optimization control strategies based on deep reinforcement learning models. While ensuring indoor environmental comfort and air quality, it can significantly reduce the overall building energy consumption and has online self-learning capabilities to continuously adapt to changes in the environment and usage patterns.

[0005] The first aspect of this invention provides a building energy-saving adaptive control method based on scenario analysis, comprising the following steps: Environmental data and equipment operating status data are acquired through a multimodal sensor network installed inside the building, and time characteristic information is set; the time characteristic information, environmental data, and equipment operating status data constitute a state space representing the building scene; Discretize or parameterize the executable control operations of the building's energy-consuming equipment to be controlled, forming an action space; set up a multi-objective reward function based on energy efficiency control objectives, comfort control objectives, and air quality control objectives; Set up the DQN model to be trained; take the state space as input, the action space as output, and the multi-objective reward function as the optimization objective to train the DQN model and obtain the pre-trained DQN model. Acquire environmental data and equipment operating status data of the building equipment to be adjusted, construct the time feature information of the current moment, and generate the current state space; The current state space is input into the pre-trained DQN model, and the DQN model outputs device control actions; a device adjustment strategy is generated based on the device control actions, and the device adjustment strategy is executed. Acquire the new status and actual energy consumption data of building equipment after the execution of the equipment adjustment strategy; calculate the immediate reward based on the multi-objective reward function; store the immediate reward, new status and actual energy consumption data in the experience replay pool of the DQN model, and adjust the DQN model through periodic sampling to update the action space of the DQN model.

[0006] Furthermore, the environmental data includes indoor temperature, indoor humidity, carbon dioxide concentration, light intensity, and occupancy information, as well as outdoor meteorological data obtained through an IoT gateway; The equipment operating status data includes real-time power and energy consumption data of the main energy-consuming equipment, as well as the set parameters of the energy-consuming equipment; The time feature information includes the current time, date type, season information, and sunshine duration indicator.

[0007] Furthermore, the discretization or parameterization of the executable control operations of the building energy-consuming equipment to be controlled, forming an action space, includes: For air conditioning systems, the adjustment action of the temperature setpoint is discretized into raising it by one degree, lowering it by one degree, or keeping it unchanged; For ventilation fans, the speed control is discretized into three levels: high, medium, and low. For lighting systems, the brightness adjustment parameter is parameterized to allow for continuous adjustment within the range of zero to one hundred percent; A binary control command for starting and stopping the fresh air system.

[0008] Furthermore, the step of setting a multi-objective reward function based on energy efficiency control targets, comfort control targets, and air quality control targets includes: Set up a composite reward function consisting of weighted summation, which mainly includes energy efficiency reward, comfort penalty and air quality penalty; The value of the energy efficiency bonus item is inversely proportional to the system's real-time total energy consumption; The value of the comfort penalty increases according to the degree to which the actual indoor temperature deviates from the range of human comfort temperature; The air quality penalty item provides a negative reward when the carbon dioxide concentration exceeds a preset health threshold.

[0009] Furthermore, the step of training the DQN model by taking the state space as input, the action space as output, and the multi-objective reward function as the optimization objective to obtain a pre-trained DQN model includes the following steps: Randomly initialize an evaluation neural network and a target neural network with the same structure; In simulated environments or historical data, the agent is allowed to choose control actions based on the current environmental state, either through random exploration with a certain probability or by using the optimal decision-making strategy of the current model. After an action is performed, the environment generates a new state and calculates an immediate reward value, which is used as interactive experience; the interactive experience is stored in a fixed-capacity experience replay pool. During training iterations, a batch of past experiences is randomly drawn from the experience pool periodically to calculate the prediction error of the neural network. The prediction error is minimized using the gradient descent algorithm, and the parameters of the evaluation network are updated. The parameters of the evaluation network are then synchronized to the target network at fixed intervals, and this process is repeated until convergence.

[0010] Furthermore, the device adjustment strategy includes: Based on the specific action instructions output by the intelligent decision, a sequence of device control commands that can be issued is generated. The lighting brightness is automatically reduced based on the light sensor readings to make full use of natural light; Based on the indoor-outdoor temperature difference and carbon dioxide concentration, the system intelligently decides to adopt different ventilation modes.

[0011] Furthermore, the step of storing instantaneous rewards, new states, and actual energy consumption data into the experience replay pool of the DQN model, and adjusting the DQN model through periodic sampling to update the action space of the DQN model includes the following steps: The new experience data generated from each decision-making interaction, including the current state, the action performed, the reward obtained, and the next state, are stored in the experience replay pool. A small batch of sample data is randomly drawn from the experience pool, the loss predicted by the model is calculated, and the network parameters of the DQN model are updated through the backpropagation algorithm.

[0012] The second aspect of the present invention provides a building energy-saving adaptive control system based on scene analysis, including a first processing module, used to acquire environmental data and equipment operating status data through a multimodal sensor network set inside the building, and set time feature information; and to construct a state space characterizing the building scene through the time feature information, environmental data and equipment operating status data. The second processing module is used to discretize or parameterize the executable control operations of the building energy-consuming equipment to be controlled, forming an action space; and to set a multi-objective reward function based on energy efficiency control objectives, comfort control objectives, and air quality control objectives. The third processing module is used to set up the DQN model to be trained; the state space is used as input, the action space is used as output, and the multi-objective reward function is used as the optimization objective to train the DQN model and obtain the pre-trained DQN model. The fourth processing module is used to acquire environmental data and equipment operating status data of the building equipment to be adjusted, construct the time feature information of the current moment, and generate the current state space; The fifth processing module is used to input the current state space into the pre-trained DQN model, and the DQN model outputs device control actions. Generate equipment adjustment strategies based on equipment control actions, and execute the equipment adjustment strategies; The sixth processing module is used to acquire the new status and actual energy consumption data of building equipment after the execution of the equipment adjustment strategy; calculate the immediate reward according to the multi-objective reward function; store the immediate reward, new status and actual energy consumption data in the experience replay pool of the DQN model, and adjust the DQN model through periodic sampling to update the action space of the DQN model.

[0013] A third aspect of the present invention provides a computer device, comprising: Memory, transceiver, processor, and bus system; The memory is used to store programs; The processor is used to execute the program in the memory, including executing the above-described building energy-saving adaptive control method based on scenario analysis; The bus system is used to connect the memory and the processor to enable communication between the memory and the processor.

[0014] A fourth aspect of the present invention provides a readable storage medium storing computer-readable instructions, which, when executed by a processor, implement the steps of the above-described building energy-saving adaptive control method based on scenario analysis.

[0015] As can be seen from the above technical solutions, the present invention has the following advantages: This invention, through deep integration of real-time multi-source sensing data and reinforcement learning decision-making, can dynamically capture the complex relationships between personnel activities, environmental changes, and equipment status, and automatically optimize and generate a global energy-saving control strategy for the current scenario. Compared to traditional control systems based on fixed schedules or threshold triggers, this invention can achieve more refined on-demand energy supply, avoiding energy waste such as high energy consumption during unattended operation or over-supply.

[0016] Secondly, traditional control systems often struggle to simultaneously address multiple environmental objectives and suffer from limited control precision. This invention, by designing a multi-objective reward function, embeds thermal comfort and air quality as core optimization objectives into the learning process. This drives the system to actively maintain environmental quality while saving energy, and continuously stabilizes air quality indicators such as CO2 concentration below healthy thresholds, thereby improving the comfort of indoor occupants.

[0017] Finally, by optimizing equipment operation strategies through reinforcement learning, this invention can smooth equipment control commands, reduce unnecessary start-stop cycles, and enable the equipment to operate more in high-efficiency operating ranges, effectively reducing the mechanical and electrical stress of the equipment, thereby extending the service life of critical energy-consuming equipment and reducing maintenance frequency and replacement costs.

[0018] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from an examination of the following, or may be learned from the practice of the invention. Attached Figure Description

[0019] Figure 1 The method flowchart provided by the present invention. Detailed Implementation

[0020] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “corresponding to,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] Example 1 The implementation method in this embodiment can be implemented in a system, on a server, or on a terminal; no specific limitation is made. The method in this application will be described from the perspective of system implementation below. As shown in the figure, a building energy-saving adaptive control method based on scenario analysis includes the following steps: Environmental data and equipment operating status data are acquired through a multimodal sensor network installed inside the building, and time characteristic information is set; the time characteristic information, environmental data, and equipment operating status data constitute a state space representing the building scene; Discretize or parameterize the executable control operations of the building's energy-consuming equipment to be controlled, forming an action space; set up a multi-objective reward function based on energy efficiency control objectives, comfort control objectives, and air quality control objectives; Set up the DQN model to be trained; take the state space as input, the action space as output, and the multi-objective reward function as the optimization objective to train the DQN model and obtain the pre-trained DQN model. Acquire environmental data and equipment operating status data of the building equipment to be adjusted, construct the time feature information of the current moment, and generate the current state space; Input the current state space into the pre-trained DQN model, and the DQN model outputs the device control actions; Generate equipment adjustment strategies based on equipment control actions, and execute the equipment adjustment strategies; Acquire the new status and actual energy consumption data of building equipment after the execution of the equipment adjustment strategy; calculate the immediate reward based on the multi-objective reward function; store the immediate reward, new status and actual energy consumption data in the experience replay pool of the DQN model, and adjust the DQN model through periodic sampling to update the action space of the DQN model.

[0022] The system collects real-time physical data reflecting the building's internal environment through a multimodal sensor network deployed within the building, including temperature and humidity sensors, CO2 sensors, light sensors, and cameras. Simultaneously, it acquires real-time operating status data for major energy-consuming equipment such as air conditioners, lighting, and fans through equipment monitoring units, including equipment settings, power consumption, and energy usage. Furthermore, the system incorporates time-related information, such as the current time, weekday / holiday, and season. For example, in an office area on a weekday afternoon, the status space might include: indoor temperature 26℃, CO2 concentration 850ppm, light intensity 300lux, air conditioner set temperature 25℃, and time 2:00 PM (weekday).

[0023] The action space refers to the set of all control operations that the system can execute, such as raising / lowering the air conditioner temperature by 1°C, setting the fan to low / medium / high speed, or adjusting the lighting brightness from 30% to 50%. The multi-objective reward function is the evaluation standard for system behavior, considering three main objectives: energy efficiency, comfort, and air quality. For example, the reward function provides positive incentives for reduced energy consumption and negative penalties for indoor temperatures deviating from the comfort range or CO2 concentrations exceeding the standard, thereby guiding the system to learn strategies that achieve a balance among multiple objectives.

[0024] This invention employs a Deep Q-Network (DQN) as its core learning algorithm. During training, the system uses historical operating data or simulation environments, taking the state space as input and the action space as output, and optimizing a multi-objective reward function. Through repeated trial and error and experience accumulation, the model gradually learns which control actions to take in different scenarios to maximize long-term cumulative rewards. After training, a pre-trained model with preliminary decision-making capabilities is obtained.

[0025] During actual system operation, the system acquires the latest environmental and equipment data in real time, constructs the current state space, and inputs it into a pre-trained DQN model. The model outputs an optimal equipment control action based on the current state, such as "lowering the air conditioner set temperature by 1°C and turning on medium-speed ventilation." The system then generates and executes specific equipment adjustment strategies to achieve real-time control of the building environment. After executing the control strategy, the system continues to monitor the new states of the environment and equipment, as well as actual energy consumption data, and calculates the immediate reward obtained for this action based on the reward function. The complete experience of this decision is stored in an experience replay pool. The system periodically extracts a batch of historical experiences from the pool to fine-tune and update the DQN model, enabling the model to continuously adapt to dynamic factors such as building usage patterns and seasonal changes, achieving continuous strategy optimization. This allows the system to not only statically execute strategies but also possess dynamic self-learning and adaptive capabilities.

[0026] Example 2 The difference between this embodiment and Embodiment 1 is that the environmental data includes indoor temperature, indoor humidity, carbon dioxide concentration, light intensity, and information on the presence of people, as well as outdoor meteorological data obtained through an IoT gateway; The equipment operating status data includes real-time power and energy consumption data of the main energy-consuming equipment, as well as the set parameters of the energy-consuming equipment; The time feature information includes the current time, date type, season information, and sunshine duration indicator.

[0027] The environmental data includes: indoor temperature and humidity reflecting the indoor thermal and humidity environment; carbon dioxide concentration reflecting air quality; light intensity reflecting natural lighting conditions; and personnel presence information used to determine personnel activity. In addition, the system connects to external meteorological data sources via an IoT gateway to obtain outdoor meteorological data such as outdoor temperature, humidity, wind speed, and solar radiation.

[0028] The equipment operation status data includes: real-time power and cumulative energy consumption data of major energy-consuming equipment such as air conditioners, lighting, fans, and water pumps, collected through smart meters or embedded sensors; at the same time, the current setting parameters of each device are obtained, such as the set temperature and operating mode of the air conditioner, the speed level of the fan, and the brightness level of the lighting circuit, so as to accurately grasp the control status of the equipment.

[0029] The time-related information includes: the current time, which reflects the daily cycle and can be accurate to the hour or minute; the date type, which distinguishes different work and rest patterns and energy consumption patterns, such as weekdays, weekends, and statutory holidays; seasonal information, which correlates climate with building thermal characteristics; and a sunshine duration marker, which identifies the time periods during the day when natural light can be utilized.

[0030] Example 3 The difference between this embodiment and embodiment two is that the executable control operations of the building energy-consuming equipment to be controlled are discretized or parameterized to form an action space, including: For air conditioning systems, the adjustment action of the temperature setpoint is discretized into raising it by one degree, lowering it by one degree, or keeping it unchanged; For ventilation fans, the speed control is discretized into three levels: high, medium, and low. For lighting systems, the brightness adjustment parameter is parameterized to allow for continuous adjustment within the range of zero to one hundred percent; A binary control command for starting and stopping the fresh air system.

[0031] For air conditioning systems, temperature adjustment actions can be discretized into multiple basic operations, such as "raise by one degree", "lower by one degree" or "keep unchanged". The adjustment accuracy can be adjusted according to the specific adjustment accuracy supported by the equipment or based on the historical data of the equipment.

[0032] For ventilation fans, their speed control can be discretized into three control levels: "high", "medium" and "low", to correspond to the common multi-level control modes of fans. The system can flexibly select different levels according to fresh air demand, indoor and outdoor pressure difference or CO2 concentration, so as to achieve energy saving while meeting ventilation requirements.

[0033] For lighting systems, a parametric approach can be used, defining brightness control as a continuously adjustable variable within the range of 0% to 100%. The decision model can output a specific percentage value, and the actuator can achieve stepless adjustment via a dimming driver, for example, adjusting to 65%.

[0034] For fresh air systems, the on / off state control can be defined as a binary control command, that is, a simple on or off. For example, during transitional seasons or at night, when natural ventilation is sufficient or there are few people, the system can output a shut-off command to save fan energy consumption.

[0035] Example 4 The difference between this embodiment and Embodiment 3 is that the step of setting a multi-objective reward function based on energy efficiency control targets, comfort control targets, and air quality control targets includes: Set up a composite reward function consisting of weighted summation, which mainly includes energy efficiency reward, comfort penalty and air quality penalty; The value of the energy efficiency bonus item is inversely proportional to the system's real-time total energy consumption; The value of the comfort penalty increases according to the degree to which the actual indoor temperature deviates from the range of human comfort temperature; The air quality penalty item provides a negative reward when the carbon dioxide concentration exceeds a preset health threshold.

[0036] The energy efficiency reward item is used to directly incentivize control behaviors that achieve energy savings. This item's value is inversely proportional to the system's real-time total energy consumption; that is, if the building's total energy consumption decreases after executing a certain control action, this item contributes a positive reward; conversely, it generates a negative reward. This guides the model to actively explore various energy-saving strategies, such as making full use of natural lighting, reducing unnecessary equipment start-ups and shutdowns, and optimizing equipment operating parameters.

[0037] The comfort penalty term is used to maintain a stable and suitable indoor thermal environment. For example, the deviation between the actual indoor temperature and the preset human comfort temperature range is used as the evaluation criterion, such as 24℃±2. The greater the deviation, the greater the negative penalty generated by this term, which means that when the model adjusts equipment such as air conditioners, it must take human thermal perception as an important constraint to avoid excessively sacrificing environmental comfort in pursuit of energy saving.

[0038] Air quality penalty is used to ensure the health needs of indoor occupants. Taking carbon dioxide concentration as an example, a health threshold is set. When the monitored CO2 concentration does not exceed this threshold, this contribution is zero; once the concentration exceeds the standard, the negative penalty increases linearly according to the magnitude of the excess, forcing the model to focus on ventilation and air exchange. When dense crowds and CO2 accumulation are detected, even if turning on fresh air or increasing fan speed will result in additional energy consumption, the model may prioritize actions to improve air quality.

[0039] By constructing a reward function, the consequences of each control decision made by the system can be quantitatively evaluated. For example, even if an action slightly increases energy consumption, it may significantly improve temperature uniformity without affecting air quality, and its overall reward may still be positive. This allows the deep reinforcement learning model to automatically learn how to find the dynamic optimal balance between the three goals of energy saving, comfort, and health through long-term training, thereby achieving intelligent building energy-saving control.

[0040] Example 5 The difference between this embodiment and embodiment four is that the step of training the DQN model by using the state space as input, the action space as output, and the multi-objective reward function as the optimization objective to obtain a pre-trained DQN model includes the following steps: Randomly initialize an evaluation neural network and a target neural network with the same structure; In simulated environments or historical data, the agent is allowed to choose control actions based on the current environmental state, either through random exploration with a certain probability or by using the optimal decision-making strategy of the current model. After an action is performed, the environment generates a new state and calculates an immediate reward value, which is used as interactive experience; the interactive experience is stored in a fixed-capacity experience replay pool. During training iterations, a batch of past experiences is randomly drawn from the experience pool periodically to calculate the prediction error of the neural network. The prediction error is minimized using the gradient descent algorithm, and the parameters of the evaluation network are updated. The parameters of the evaluation network are then synchronized to the target network at fixed intervals, and this process is repeated until convergence.

[0041] First, the neural network is initialized. A dual-network structure can be used to ensure training stability, including an evaluation network and a target network, both with the same network architecture; for example, a multi-layer fully connected neural network. At the start of training, the weight parameters of the evaluation network are randomly initialized, and the same parameters are copied to the target network.

[0042] Secondly, interactive exploration and learning are conducted in a simulation environment or on historical datasets. During training, the DQN model to be trained selects actions based on the currently observed environmental state according to the ε-greedy policy. The ε-greedy policy explores randomly with a certain probability, trying actions that have not been selected before to avoid getting trapped in local optima; at the same time, it uses the remaining probability to utilize the actions that the current model considers optimal to obtain immediate rewards. For example, in the early stages of training, the system may conduct more random exploration to collect feedback from different state-action pairs.

[0043] Then, interaction experiences are formed and stored. After the agent executes a selected action, the environment is updated to a new state, and the immediate reward for that action is calculated according to the reward function. This constitutes a complete interaction experience: current state, action executed, reward obtained, and next state. The corresponding experience is stored in an experience replay pool with a fixed capacity. The design of the experience pool allows the system to reuse historical data, break the temporal correlation between data, improve sample efficiency, and stabilize the training process.

[0044] Next, during the training iterations, the system periodically and randomly extracts a small batch of historical experience data from the experience replay pool. Using this data, the prediction accuracy of the current evaluation network is evaluated by calculating the temporal difference error. That is, the target network is used to calculate the expected maximum future reward of the next state, which, combined with the immediate reward, constitutes the target value for training; this is compared with the evaluation network's prediction of the current state-action pair, and the difference is the prediction error.

[0045] Finally, a loss function is constructed using the calculated prediction error, and the gradient descent algorithm is used to update the weight parameters of the evaluation network to reduce the prediction error and make the model's Q-value estimation more accurate. Simultaneously, to maintain the stability of the training objective, the parameters of the evaluation network are synchronously copied to the target network at fixed training steps. This process is iterated until the model performance stabilizes, which is considered training convergence, resulting in a pre-trained DQN model that can be used for online decision-making.

[0046] Example 6 The difference between this embodiment and embodiment five is that the device adjustment strategy includes: Based on the specific action instructions output by the intelligent decision, a sequence of device control commands that can be issued is generated. The lighting brightness is automatically reduced based on the light sensor readings to make full use of natural light; Based on the indoor-outdoor temperature difference and carbon dioxide concentration, the system intelligently decides to adopt different ventilation modes.

[0047] By receiving device control actions output from the DQN model, and according to the device communication protocol and interface specifications, these actions are encapsulated into standard control commands and dispatched to the corresponding building automation system or IoT gateway for execution. The strategy includes automated fine-tuning rules based on environmental awareness. To improve system response speed and reliability, the strategy layer can incorporate fast-response conditional rules, complementing the decision-making of the DQN model.

[0048] For example, a strategy can automatically reduce or turn off electrical lighting in a corresponding area when there is sufficient natural light, based on real-time readings from a light sensor, thereby achieving energy savings directly and quickly without having to go through model inference every time.

[0049] For complex scenarios involving the interaction of multiple subsystems, the strategy layer defines specific collaborative logic. For example, based on the indoor-outdoor temperature difference and CO2 concentration, the system can intelligently decide and execute different ventilation modes.

[0050] For example, when the outdoor temperature is suitable and lower than the indoor temperature, and the CO2 concentration does not exceed the standard, the strategy can be to implement natural ventilation mode, that is, open the windows or exhaust windows and turn off the refrigeration unit; when the temperature difference between indoor and outdoor is small but the CO2 concentration is high, mechanical ventilation mode is activated, that is, the fresh air unit is run to exchange air; when the outdoor temperature is extremely high or extremely low, return air mixing mode is adopted, which circulates air and handles temperature and humidity on the basis of minimum fresh air volume.

[0051] Example 7 The difference between this embodiment and embodiment six is ​​that the step of storing the immediate reward, new state, and actual energy consumption data into the experience replay pool of the DQN model, and adjusting the DQN model through periodic sampling to update the action space of the DQN model includes the following steps: The new experience data generated from each decision-making interaction, including the current state, the action performed, the reward obtained, and the next state, are stored in the experience replay pool. A small batch of sample data is randomly drawn from the experience pool, the loss predicted by the model is calculated, and the network parameters of the DQN model are updated through the backpropagation algorithm.

[0052] During operation, in each online control cycle, after the system executes the device adjustment strategy based on the current state, it simultaneously monitors environmental feedback and the new state, and calculates the immediate reward. Subsequently, the current state formed by this interaction, the action executed, the immediate reward obtained, and the new state observed after the action are executed are stored as a data sample in a fixed-capacity experience replay pool with first-in-first-out characteristics.

[0053] Secondly, the system initiates a model update batch by setting a trigger cycle, such as at regular time intervals or after accumulating a certain amount of new experience. In this batch, the system randomly selects a small batch of samples from the experience replay pool.

[0054] Next, the system uses the extracted empirical data to perform forward propagation and loss calculation. For example, for each sample, the evaluation network calculates the predicted value of the action in the current state; simultaneously, the target network calculates the maximum expected value of all possible actions in the next state, and combines this with the immediate reward and discount factor to form the target value. The mean squared error between the predicted value and the target value constitutes the loss function. The system then uses the backpropagation algorithm to calculate the gradient of the loss function with respect to the parameters of the evaluation network, and uses the gradient descent optimizer to perform a small update on the weight parameters of the evaluation network. Example 8 A building energy-saving adaptive control system based on scenario analysis includes a first processing module, which is used to acquire environmental data and equipment operating status data through a multimodal sensor network set up inside the building, and set time feature information; and construct a state space representing the building scenario through time feature information, environmental data and equipment operating status data. The second processing module is used to discretize or parameterize the executable control operations of the building energy-consuming equipment to be controlled, forming an action space; and to set a multi-objective reward function based on energy efficiency control objectives, comfort control objectives, and air quality control objectives. The third processing module is used to set up the DQN model to be trained; the state space is used as input, the action space is used as output, and the multi-objective reward function is used as the optimization objective to train the DQN model and obtain the pre-trained DQN model. The fourth processing module is used to acquire environmental data and equipment operating status data of the building equipment to be adjusted, construct the time feature information of the current moment, and generate the current state space; The fifth processing module is used to input the current state space into the pre-trained DQN model, and the DQN model outputs device control actions. Generate equipment adjustment strategies based on equipment control actions, and execute the equipment adjustment strategies; The sixth processing module is used to acquire the new status and actual energy consumption data of building equipment after the execution of the equipment adjustment strategy; calculate the immediate reward according to the multi-objective reward function; store the immediate reward, new status and actual energy consumption data in the experience replay pool of the DQN model, and adjust the DQN model through periodic sampling to update the action space of the DQN model.

[0055] Example 9 A computer device, comprising: Memory, transceiver, processor, and bus system; The memory is used to store programs; The processor is used to execute the program in the memory, including executing the above-described building energy-saving adaptive control method based on scenario analysis; The bus system is used to connect the memory and the processor to enable communication between the memory and the processor.

[0056] Example 10 A readable storage medium storing computer-readable instructions, which, when executed by a processor, implement the steps of the above-described building energy-saving adaptive control method based on scenario analysis.

[0057] In summary, this invention, through deep integration of real-time multi-source sensing data and reinforcement learning decision-making, can dynamically capture the complex relationships between personnel activities, environmental changes, and equipment status, and automatically optimize and generate a global energy-saving control strategy for the current scenario. Compared to traditional control systems based on fixed schedules or threshold triggers, this invention can achieve more refined on-demand energy supply, avoiding energy waste such as high energy consumption during unattended operation or over-supply.

[0058] Secondly, traditional control systems often struggle to simultaneously address multiple environmental objectives and suffer from limited control precision. This invention, by designing a multi-objective reward function, embeds thermal comfort and air quality as core optimization objectives into the learning process. This drives the system to actively maintain environmental quality while saving energy, and continuously stabilizes air quality indicators such as CO2 concentration below healthy thresholds, thereby improving the comfort of indoor occupants.

[0059] Finally, by optimizing equipment operation strategies through reinforcement learning, this invention can smooth equipment control commands, reduce unnecessary start-stop cycles, and enable the equipment to operate more in high-efficiency operating ranges, effectively reducing the mechanical and electrical stress of the equipment, thereby extending the service life of critical energy-consuming equipment and reducing maintenance frequency and replacement costs.

[0060] Those skilled in the art will recognize that the units of the various examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the invention.

[0061] In the embodiments provided by the present invention, it should be understood that the division of units is only a logical functional division. In actual implementation, there may be other division methods, such as multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored.

[0062] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0063] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0064] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A building energy-saving adaptive control method based on scenario analysis, characterized in that, Includes the following steps: Environmental data and equipment operating status data are acquired through a multimodal sensor network installed inside the building, and time characteristic information is set; the time characteristic information, environmental data, and equipment operating status data constitute a state space representing the building scene; Discretize or parameterize the executable control operations of the building's energy-consuming equipment to be controlled, thus forming an action space; A multi-objective reward function is set based on energy efficiency control targets, comfort control targets, and air quality control targets; Set up the DQN model to be trained; take the state space as input, the action space as output, and the multi-objective reward function as the optimization objective to train the DQN model and obtain the pre-trained DQN model. Acquire environmental data and equipment operating status data of the building equipment to be adjusted, construct the time feature information of the current moment, and generate the current state space; Input the current state space into the pre-trained DQN model, and the DQN model outputs the device control actions; Generate equipment adjustment strategies based on equipment control actions, and execute the equipment adjustment strategies; Acquire the new status and actual energy consumption data of building equipment after the equipment adjustment strategy is executed; Calculate immediate rewards based on a multi-objective reward function; Real-time rewards, new states, and actual energy consumption data are stored in the experience replay pool of the DQN model. The DQN model is adjusted and its action space is updated through periodic sampling.

2. The building energy-saving adaptive control method based on scenario analysis according to claim 1, characterized in that, The environmental data includes indoor temperature, indoor humidity, carbon dioxide concentration, light intensity, and information on the presence of people, as well as outdoor meteorological data obtained through an IoT gateway; The equipment operating status data includes real-time power and energy consumption data of the main energy-consuming equipment, as well as the set parameters of the energy-consuming equipment; The time feature information includes the current time, date type, season information, and sunshine duration indicator.

3. The building energy-saving adaptive control method based on scenario analysis according to claim 1, characterized in that, The process of discretizing or parameterizing the executable control operations of the building energy-consuming equipment to be controlled, forming an action space, includes: For air conditioning systems, the adjustment action of the temperature setpoint is discretized into raising it by one degree, lowering it by one degree, or keeping it unchanged; For ventilation fans, the speed control is discretized into three levels: high, medium, and low. For lighting systems, the brightness adjustment parameter is parameterized to allow for continuous adjustment within the range of zero to one hundred percent; A binary control command for starting and stopping the fresh air system.

4. The building energy-saving adaptive control method based on scenario analysis according to claim 1, characterized in that, The step of setting a multi-objective reward function based on energy efficiency control targets, comfort control targets, and air quality control targets includes: Set up a composite reward function consisting of weighted summation, which mainly includes energy efficiency reward, comfort penalty and air quality penalty; The value of the energy efficiency bonus item is inversely proportional to the system's real-time total energy consumption; The value of the comfort penalty increases according to the degree to which the actual indoor temperature deviates from the range of human comfort temperature; The air quality penalty item provides a negative reward when the carbon dioxide concentration exceeds a preset health threshold.

5. The building energy-saving adaptive control method based on scenario analysis according to claim 1, characterized in that, The process of training a DQN model by using the state space as input, the action space as output, and a multi-objective reward function as the optimization objective, to obtain a pre-trained DQN model, includes the following steps: Randomly initialize an evaluation neural network and a target neural network with the same structure; In simulated environments or historical data, the agent is allowed to choose control actions based on the current environmental state, either through random exploration with a certain probability or by using the optimal decision-making strategy of the current model. After an action is performed, the environment generates a new state and calculates an immediate reward value, which is used as interactive experience; the interactive experience is stored in a fixed-capacity experience replay pool. During training iterations, a batch of past experiences is randomly drawn from the experience pool periodically to calculate the prediction error of the neural network. The prediction error is minimized using the gradient descent algorithm, and the parameters of the evaluation network are updated. The parameters of the evaluation network are then synchronized to the target network at fixed intervals, and this process is repeated until convergence.

6. The building energy-saving adaptive control method based on scenario analysis according to claim 1, characterized in that, The equipment adjustment strategy includes: Based on the specific action instructions output by the intelligent decision, a sequence of device control commands that can be issued is generated. The lighting brightness is automatically reduced based on the light sensor readings to make full use of natural light; Based on the indoor-outdoor temperature difference and carbon dioxide concentration, the system intelligently decides to adopt different ventilation modes.

7. The building energy-saving adaptive control method based on scenario analysis according to claim 1, characterized in that, The process of storing instant rewards, new states, and actual energy consumption data into the experience replay pool of the DQN model, and adjusting the DQN model through periodic sampling to update the action space of the DQN model, includes the following steps: The new experience data generated from each decision-making interaction, including the current state, the action performed, the reward obtained, and the next state, are stored in the experience replay pool. A small batch of sample data is randomly drawn from the experience pool, the loss predicted by the model is calculated, and the network parameters of the DQN model are updated through the backpropagation algorithm.

8. A building energy-saving adaptive control system based on scenario analysis, characterized in that, It includes a first processing module, which is used to acquire environmental data and equipment operating status data through a multimodal sensor network set up inside the building, and set time feature information; and to construct a state space representing the building scene through time feature information, environmental data and equipment operating status data; The second processing module is used to discretize or parameterize the executable control operations of the building energy-consuming equipment to be controlled, forming an action space. A multi-objective reward function is set based on energy efficiency control targets, comfort control targets, and air quality control targets; The third processing module is used to set up the DQN model to be trained; the state space is used as input, the action space is used as output, and the multi-objective reward function is used as the optimization objective to train the DQN model and obtain the pre-trained DQN model. The fourth processing module is used to acquire environmental data and equipment operating status data of the building equipment to be adjusted, construct the time feature information of the current moment, and generate the current state space; The fifth processing module is used to input the current state space into the pre-trained DQN model, and the DQN model outputs device control actions. Generate equipment adjustment strategies based on equipment control actions, and execute the equipment adjustment strategies; The sixth processing module is used to acquire the new status and actual energy consumption data of building equipment after the execution of the equipment adjustment strategy; and to calculate the real-time reward based on the multi-objective reward function. Real-time rewards, new states, and actual energy consumption data are stored in the experience replay pool of the DQN model. The DQN model is adjusted and its action space is updated through periodic sampling.

9. A computer device, characterized in that, include: Memory, transceiver, processor, and bus system; The memory is used to store programs; The processor is used to execute the program in the memory, including executing a building energy-saving adaptive control method based on scenario analysis as described in any one of claims 1 to 7; The bus system is used to connect the memory and the processor to enable communication between the memory and the processor.

10. A readable storage medium storing computer-readable instructions, characterized in that, When the computer-readable instructions are executed by the processor, they implement the steps of a building energy-saving adaptive control method based on scenario analysis as described in any one of claims 1 to 7.

Images