Central air conditioner load self-adaptive regulation and control method based on artificial intelligence

Abstract

The invention discloses a central air conditioner load self-adaptive regulation and control method based on artificial intelligence, which relates to the technical field of heating ventilation air conditioner automatic regulation and control, and comprises the steps of edge end data preprocessing, deep learning load prediction and self-adaptive updating, intelligent control strategy, closed-loop feedback and continuous optimization and fault diagnosis and remote operation and maintenance. Through edge end denoising and feature labeling, the data transmission pressure is remarkably reduced, and the response speed is increased; the deep learning model is combined with online iteration and reinforcement learning, key parameters such as chilled water temperature and water pump frequency are accurately regulated and controlled, and energy consumption and comfort are considered; according to the method, prediction errors and equipment states are monitored, a model and a control strategy are synchronously updated when abnormity is detected, fault early warning is triggered in combination with health assessment, energy consumption is effectively reduced, the service life of equipment is prolonged, and finally an efficient and stable self-adaptive regulation and control system capable of being operated and maintained online in a long period is constructed.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic regulation of heating, ventilation and air conditioning, and specifically to an adaptive regulation method for central air-conditioning load based on artificial intelligence. Background Art

[0002] In modern large public buildings and complexes, the central air-conditioning system accounts for a considerable proportion of the overall energy consumption and directly affects the comfort and health of the indoor environment. In such application scenarios, multi-source variables such as indoor and outdoor temperature, humidity, personnel flow, and solar radiation intensity show great volatility with the change of seasons and time periods; at the same time, the operation modes of various terminal devices (including chilled water units, water pumps, fans, etc.) are also driven by complex load demands. In order to balance energy conservation and environmental comfort, the operator usually needs to finely manage the energy consumption curve, equipment operating conditions, and building function usage, but the traditional timing or empirical regulation methods often lack a forward-looking assessment of real-time load fluctuations and equipment health conditions, resulting in poor energy efficiency or the risk of unexpected shutdowns, especially in scenarios with high personnel density or frequent heat source disturbances.

[0003] In the Chinese invention patent with the application publication number CN108036464A, an adaptive dynamic cooling load regulation method for a central air-conditioning system is provided, which is used in the central air-conditioning system to achieve volume regulation according to the change of dynamic cooling load; this method realizes variable temperature difference operation in the way of actively changing the pressure difference and flow rate according to the outdoor temperature. The cooling load of the central air-conditioning system is divided into two parts, one part is the steady-state cooling load related to the outdoor temperature, and the other part is the uncertain dynamic cooling load caused by hot water, solar radiation and indoor free heat. The variable temperature difference in the way of actively changing the pressure difference and flow rate follows the change of the steady-state cooling load; the passive temperature difference change follows the dynamic cooling load, which can effectively meet the demand of cooling on demand and realize the power-saving operation of the central air-conditioning.

[0004] However, combining the above actual application scenarios and the content in the prior art: To achieve precise control of the central air-conditioning system in such a dynamic and uncertain environment, the core problem lies in how to efficiently integrate multi-source data from distributed sensors and device terminals, and based on the analysis and modeling of large-scale heterogeneous data streams, real-time predict the air-conditioning load demand and equipment operating status, and then perform adaptive scheduling on the chilled water supply temperature, water pump frequency, valve opening, etc.; at the same time, it is also necessary to establish a relative balance among sensor noise, equipment fault hazards, energy consumption costs and environmental comfort during the long-term operation of the system.

[0005] The prior art either adopts passive and single-loop control means, making it difficult to respond to drastic load fluctuations in a timely manner, or simply stays at the level of adjusting empirical parameters, and cannot provide a systematic solution that takes into account both energy consumption optimization and equipment health management for building operation and maintenance parties, bringing huge technical challenges to building energy conservation and fault prevention.

[0006] For this reason, the present invention provides an adaptive control method for central air-conditioning load based on artificial intelligence. Summary of the Invention

[0007] (I) Technical problems to be solved Aiming at the deficiencies of the prior art, the present invention provides an adaptive control method for central air-conditioning load based on artificial intelligence. Through edge denoising and feature annotation, the data transmission pressure is significantly reduced and the response speed is improved; the deep learning model combines online iteration and reinforcement learning to accurately control key parameters such as chilled water temperature and pump frequency, taking into account both energy consumption and comfort; for prediction error and equipment status monitoring, when an anomaly is detected, the model and control strategy are updated synchronously, and a fault warning is triggered in combination with health assessment, effectively reducing energy consumption and extending the equipment life. Finally, an adaptive control system that is efficient, stable and can be operated online for a long period is constructed to solve the technical problems recorded in the background art.

[0008] (II) Technical solutions To achieve the above objectives, the present invention is realized through the following technical solutions: An adaptive control method for central air-conditioning load based on artificial intelligence, including that when each distributed sensor completes the acquisition of original data such as indoor and outdoor temperature and humidity, wind speed and main equipment power within a predetermined sampling period and sends it to the edge computing node, denoising filtering and multi-dimensional feature annotation are performed on the above data objects, and a multi-source feature set with normalization and credibility identification is output; When the edge computing node generates a feature vector containing key features of the environment and equipment and transmits it to the cloud, a deep learning model is used to train and update the equipment data online, and a model output that can accurately predict the air-conditioning load demand is produced in real time and provides a reliable prediction basis for the reinforcement learning control strategy; When the load prediction value is synchronized with the multi-source state information at the current moment and meets the control cycle requirements, dynamic decision-making is performed on control variables such as chilled water temperature, pump frequency and valve opening through deep reinforcement learning, and an action vector is output on the premise of comprehensive energy consumption and comfort and sent to each sub-device; When each sub-device completes the execution according to the action vector and returns the feedback containing indoor environment, equipment power and real-time load data, with the help of the operation performance function Continuously monitor and evaluate the prediction error and control effect, and online update the parameters of the deep learning model and control algorithm when the deviation exceeds the threshold; When accumulating the equipment operation trajectory and abnormal records and detecting signs of failure or performance degradation, through the fault symptom function Jointly conduct online diagnosis and early warning on core equipment such as chilled water units and pumps with the remote operation and maintenance module, store and synchronize the diagnosis results and maintenance suggestions to the knowledge base to ensure the coordinated promotion of equipment health status and energy-saving strategies.

[0009] Preferably, deploy distributed sensors inside and outside the building, and send the raw data to the edge computing node at the time stamp The collected raw data includes: indoor temperature , outdoor relative humidity , solar irradiance , wind speed , indoor number of people and the real-time power of the main equipment of the central air conditioner (such as chilled water units, heat pumps, etc.); Preferably, initially clean the raw data, where: Let represent the measurement value of the sensor at time , let be the reference or expected central value of the sensor , and define the rolling time window length . At the current time , the cumulative deviation of the sensor is:

[0010] In the formula: is the integral window length of the sensor ; is the power coefficient; Define the abnormal metric function as follows:

[0011] In the formula: is the cumulative deviation; is the reference threshold of the cumulative deviation, is the sensitivity coefficient, ; If the cumulative deviation of the sensor significantly exceeds the reference threshold , indicating a continuous or significant deviation during this period;

[0012] Indicates the credibility; if the credibility is lower than a pre-set credibility threshold, the edge computing node will mark it as suspicious data; Preferably, normalize the data that has been cleaned and anomaly-detected, and convert the measured value into a dimensionless characteristic numerical value ; assign characteristic weights to different sensor characteristics , and pack the normalization result with its corresponding credibility and weight into a feature vector ; Preferably, use the feature vector as the model input, initialize the deep learning network structure, and initialize the weight parameters randomly or based on prior knowledge; Preferably, introduce a loss function that combines custom fusion weights and activation operations , which is used to characterize the difference between the model output and the actual load ; use the loss function to jointly optimize the network parameters through an iterative update until the stopping conditions such as accuracy or the number of iteration rounds are met; Define the following loss function :

[0013] Where: is the set of time indices corresponding to the training data; is the predicted value of the model at , is the actual load; is an adjustable weight function, which is used to combine the credibility in the feature vector and the feature weight and other information, and can be defined as:

[0014] Among them, is the sensitivity coefficient, is the number of features; is the non-linear error metric function; Preferably, when new operation feedback data is collected in real time and compared with the current model prediction result, if it is detected that the loss function of the load prediction model exceeds the preset threshold, an online learning mechanism is started: the new operation feedback data and its corresponding actual load are incorporated into a small-scale incremental training set; according to the loss function perform rapid iteration to update the model parameters; Preferably, according to the load prediction value and the current environmental state data of the building, determine the control objectives of the central air-conditioning system, including: minimizing energy consumption, maintaining the indoor temperature and humidity within the set range, and avoiding frequent start-stop or overloading operation; Establish decision variables to describe the operation states of the sub-devices of the air-conditioning system, including: is the chilled water supply temperature, is the pump operation frequency, is the chilled water valve opening degree, and is the start-stop indication of the device , where can represent key devices such as different chilled water units and fans; Introduce an objective function with adjustable weights , by minimizing this objective function, the balance between energy consumption and comfort can be achieved, and at the same time, the over-frequent start-stop of the device can be reduced; Preferably, define a state vector composed of the predicted value and the current measured environmental and device parameters, including:

[0015] Define an action vector , corresponding to ; Define a reward function to be consistent with the objective function or make a numerical transformation:

[0016] Drive the reinforcement learning policy network to continuously optimize the action vector , during the training process, continuously iterate and update the policy network parameters , so that when facing different external environments and load prediction values , it can select an approximately optimal action vector ; Preferably, after completing the decision of the real-time optimization algorithm or the reinforcement learning policy, the obtained action vector needs to be sent to each sub-device of the air-conditioning system, including: Set the target chilled water temperature of the chiller , adjust the frequency of the pump inverter , control the opening degree of the chilled water valve ; Start or stop specific chilled water units or air conditioning terminal fans ; After actual execution, the updated equipment operation data can be periodically compared with the predicted load value . If the deviation between the two is greater than expected, possible modeling errors or equipment abnormalities can be determined, and the online learning mechanism can be triggered for model adaptation; Preferably, continuously collect the real-time operation status data of the current central air conditioner; compare the real-time operation status data with the load prediction value , action vector , and introduce a custom operation performance function to measure the operation deviation at the current moment, and define the operation performance function as follows:

[0017] In the formula: is the evaluation time window length, and the system deviation and control status are cumulatively analyzed within ; is the actual load at time ; is the load prediction value of the deep learning load prediction model at time , and the model parameters are denoted as ; is the amplification coefficient of the power deviation term; is the power term; is the coefficient of the deviation change rate term; is the penalty or incentive function for the action vector ; Preferably, when it is detected that the operation performance function continuously exceeds the preset threshold for a period of time, trigger the fast iteration optimization process, including: Based on the newly collected actual load and the feature vector at the corresponding time , perform short-cycle or mini-batch updates on, and correct the prediction deviation control strategy parameters correction: Perform small-step iterations on the objective function or the parameters of the reinforcement learning policy network ; Preferably, after each trigger of fast iteration and completion of model and control strategy updates, the newly learned parameters or control strategy Store it and its corresponding operating environment information in the knowledge base; when subsequent consistent or similar scenarios occur, the similar environments or load conditions in the historical archives can be quickly matched first, and relevant experience and parameter sets can be selected as initial values; Preferably, the multi-source device status data collected is fused with the action vector and the energy consumption feedback information to form a real-time status vector for each key device; Construct a fault symptom function , and measure the difference between the current device status and the reference health baseline within the time window as follows:

[0018] In the formula: is the multi-dimensional status vector of the th device at time ; is the health baseline vector of the th device; represents the norm of the device's real-time status and the health baseline, is the power coefficient; is the size of the rolling time window for fault detection; is the order of the fractional integral; is the Gamma function, related to the definition of fractional integral, is the scaling coefficient; If the fault symptom function continuously exceeds the set fault threshold within a certain period of time, or shows high-frequency fluctuations, a fault warning is triggered and the data and analysis results are reported to the remote operation and maintenance platform; Preferably, when the fault symptom function exceeds the fault threshold and lasts for a certain duration, a fault warning event is automatically generated, and data such as the device status vector and the fault symptom trend chart are uploaded to the operation and maintenance platform; Based on the knowledge base, the operation and maintenance platform provides fault handling suggestions, including maintenance items, required spare parts, or temporary scheduling strategies; before the operation and maintenance personnel intervene, temporary control strategy optimization is carried out: fine-tuning the action vector ; (III) Beneficial effects The present invention provides an air-conditioning load adaptive regulation method based on artificial intelligence, having the following beneficial effects: By deploying distributed sensors to collect multi-source environmental and device data and performing preprocessing and anomaly screening at the edge, the data volume can be reduced. Relying on the denoising and feature annotation mechanisms, data fluctuations and noise can be accurately identified, reducing the latency caused by traditional single-node or large-scale background processing, thereby enhancing timeliness and stability. Using deep learning algorithms to construct an adaptive update load prediction model makes the capture of central air-conditioning load changes more sensitive. The input of the load prediction model includes not only basic features such as indoor and outdoor temperature, humidity, and personnel flow, but also building functions and historical energy consumption information. Combining credibility marking and online learning mechanisms, the prediction deviation is continuously iteratively corrected, improving the prediction accuracy and reducing the dependence on manual parameter tuning. Based on the prediction results Deep reinforcement learning can dynamically decide key parameters such as chilled water temperature, pump frequency, and chilled water valve opening to form a forward-looking and flexible control strategy. This process not only takes into account energy consumption costs and indoor comfort, but also effectively reduces equipment wear caused by frequent starts and stops, improving the overall energy efficiency utilization rate. By setting up a closed-loop feedback and continuous optimization process, the scheme can monitor the actual operating status in real time. By measuring the prediction error and control effect, once the deviation exceeds the threshold, the load prediction model and control strategy are adaptively updated to form a complete closed-loop management, enabling the central air-conditioning to maintain high accuracy and stability in the face of changing environments and demands. Through the integration of fault diagnosis and remote operation and maintenance, not only energy consumption optimization is concerned, but also the health status of key equipment (such as chilled water units, pumps, valves, etc.) is more comprehensively covered. Once a serious anomaly is detected through the fault symptom function an alarm can be remotely issued and the knowledge base can be linked for maintenance decision-making, which can avoid unexpected shutdowns or fault spread and improve the reliability and operation and maintenance efficiency of the central air-conditioning system throughout its life cycle. Brief Description of the Drawings

[0019] Figure 1 It is a schematic flow diagram of the method for adaptively regulating the load of the central air-conditioning according to the present invention. Detailed Embodiments

[0020] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0021] Please refer to Figure 1 , the present invention provides a method for adaptively regulating the load of a central air-conditioning based on artificial intelligence, including Step 1: When each distributed sensor completes the acquisition of original data such as indoor and outdoor temperature and humidity, wind speed, and the power of main equipment within a predetermined sampling period and sends it to the edge computing node, perform denoising filtering and multi-dimensional feature annotation on the above data objects, and output a multi-source feature set with normalization and credibility identification; The first step includes the following contents: Step 101: Distributed data acquisition and synchronization Deploy a variety of distributed sensors inside and outside the building, including indoor temperature, outdoor temperature, outdoor humidity, solar irradiance, wind speed, number of people, and the power of main air-conditioning equipment, etc.; all sensors use a unified sampling period for measurement, and at the time stamp send the original data to the edge computing node through a dedicated bus or high-bandwidth network; The collected original data includes: indoor temperature , outdoor relative humidity , solar irradiance , wind speed , number of indoor people and the real-time power of main central air-conditioning equipment (such as chilled water units, heat pumps, etc.) ; Step 102: Data cleaning and preliminary anomaly detection Perform preliminary cleaning on the original data at the edge computing node, such as removing invalid data caused by packet loss or format errors, where: Let represent the measurement value of sensor at time , let be the reference or expected central value of sensor , and define the rolling time window length . At the current time , the cumulative deviation of sensor is:

[0022] In the formula: is the integration window length of sensor ; is the power coefficient, usually taking : when it is absolute value integration, when it is square penalty, which emphasizes the amplification of deviation more; is more sensitive to extreme deviations; Compare the cumulative deviation Make a comparison and introduce the sensitivity coefficient Perform a non-linear activation mapping and define an anomaly metric function As follows:

[0023] In the formula: is the cumulative deviation; is the reference threshold of the cumulative deviation, is the sensitivity coefficient, , If the sensor 's cumulative deviation significantly exceeds the reference threshold , it indicates that there is a continuous or drastic deviation during this period;

[0024] represents the credibility; if the credibility is lower than the pre-set credibility threshold, the edge computing node will mark it as requiring further verification; When in use, by assigning credibility to each piece of data, high-reliability and low-reliability data can be treated differently in subsequent modeling or control strategies, reducing the interference caused by outliers. Compared with the traditional methods of using threshold comparison or variance judgment, introducing an adjustable sensitivity coefficient of the logistic function makes anomaly detection more flexible and scalable.

[0025] Step 103, Edge-side feature extraction and data packaging Perform normalization on the data that has completed cleaning and anomaly detection, and convert the measured value into a dimensionless feature value ; Assign feature weights to different sensor features , for example, a higher weight can be given to the number of people to more sensitively reflect the change in the flow of people in subsequent load forecasting; Package the normalization result with its corresponding credibility and weight into a feature vector ;

[0026] In the formula: is the value after normalization of the sensor ; is the credibility, is the feature weight; During use, feature extraction and packaging are completed at the edge side, which can reduce the upload volume of raw data, significantly reduce the network load, directly embed the feature weights and credibility into the feature vector, and there is no need to maintain this information separately during subsequent modeling, simplifying the management of model inputs. Normalization, credibility, and feature weights are integrated into the same data structure, effectively improving the integrity and usability of the data. Performing various preprocessing steps in advance on the edge side can better adapt to the highly dynamic building environment and win more real-time computing power for subsequent complex models and control strategies.

[0027] Step 2: When the edge computing node generates a feature vector containing the key features of the environment and equipment After transmitting it to the cloud, use a deep learning model to train and update the device data online, and generate a model output that can accurately predict the air-conditioning load demand in real time And provide a reliable prediction basis for the reinforcement learning control strategy; The above Step 2 includes the following contents: Step 201: Feature input and model initialization Use the feature vector As the model input, where each moment Contains several normalized features , corresponding credibility And feature weights ; Initialize the deep learning network structure, such as the number of layers, neurons, and activation function types of the LSTM or Transformer model, and initialize the weight parameters randomly or based on prior knowledge; The model input tensor Integrated from the feature vector Through a unified mapping function is used to adapt to the input dimension requirements of the deep learning model; Is the set of trainable parameters of the deep learning network (such as LSTM, Transformer, etc.), including the weight matrix and bias term of each layer of the loss function; Is the load prediction output of the deep learning model at time Under the parameter ; Step 202: Load prediction model training and iterative optimization Use historical data or batch data sampled in real time to train or fine-tune the model, and introduce a loss function that customizes the fusion weight and activation operation , used to characterize the difference between the model output And the actual load ; From the loss function Use joint gradient descent, Adam, or other advanced optimization algorithms to iteratively update the network parameters until the stopping conditions such as accuracy or the number of iteration rounds are met; Define the following loss function to fuse the credibility and the adaptive penalty mechanism to better handle outliers or noise interference:

[0028] where: is the set of time indices corresponding to the training data; is the predicted value of the model at , is the actual load; is an adjustable weight function used to combine the credibility in the feature vector and the feature weight and other information, and can be defined as:

[0029] where, is the sensitivity coefficient, with a value greater than 0, used to amplify or weaken the credibility accumulation effect; is the number of features; is a non - linear error metric function, which can be designed as the integral of the power of the absolute value of the prediction error or other creative metrics with calculus characteristics, such as:

[0030] where is the scaling coefficient, with a value greater than 0, is the power order, generally taking values from 1 to 3; When in use, after integrating the credibility and the non - linear metric, the loss function can automatically reduce the impact on data with large noise and give greater weights to data with higher criticality (with a higher adjustable weight function ), improving the overall training effect; Incorporate the extracted and other information into the weight part of the loss function to form a cross - step data and algorithm coupling mechanism.

[0031] Step 203, Adaptive Update and Online Learning Mechanism New operation feedback data is collected in real time, such as actual device parameters (e.g., chilled water temperature, pump frequency, valve opening, etc.), energy consumption data, and relevant environmental indicators (e.g., indoor temperature, humidity, etc.). These data are compared with the current model prediction results to evaluate the prediction accuracy of the model. If it is detected that the model prediction error exceeds the preset threshold or remains high for a certain period of time, the online learning mechanism is activated: The new operation feedback data and its corresponding actual load are incorporated into a small-scale incremental training set; According to the loss function perform rapid iteration to update the model parameters for , so that the load prediction model always maintains a high degree of matching with the latest operating conditions; During use, through the adaptive update mechanism, the model can quickly adjust its own parameters according to the latest environmental changes or building usage conditions, and continuously maintain high prediction accuracy. The micro-batch or incremental update method of online learning takes into account both real-time performance and computational overhead, ensuring dynamic correction without large-scale offline retraining.

[0032] Step 3: When the load prediction value is synchronized with the multi-source status information at the current moment and meets the control cycle requirements, perform dynamic decision-making on control variables such as chilled water temperature, pump frequency, and valve opening through deep reinforcement learning, and output an action vector under the premise of comprehensive energy consumption and comfort and send it to each sub-device; The said step 3 includes the following contents: Step 301: Unified modeling of control objectives and decision variables According to the load prediction value and the current environmental state data of the building, determine the control objectives of the central air-conditioning system, including: minimizing energy consumption, maintaining indoor temperature and humidity within the set range, and avoiding frequent start-stop or overloading operation; Establish a set of decision variables to describe the operating states of each sub-device of the air-conditioning system, including: is the chilled water supply temperature, is the pump operating frequency, is the chilled water valve opening, and is the start-stop indication (Boolean or ) of the device , where can represent different key devices such as chilled water units, fans, etc.; To take into account multi-objective requirements, introduce an adjustable-weight objective function to measure the operating state at each moment:

[0033] Wherein: represents the predicted energy consumption cost; represents the penalty for the deviation of indoor temperature and humidity from the expected range; represents the penalty or incentive for the start-stop and safe operation constraints of the equipment; are weight coefficients, all with values greater than 0, and can be dynamically adjusted according to priorities; By minimizing this objective function, the balance between energy consumption and comfort can be achieved, and at the same time, the excessive and frequent start-stop of the equipment can be reduced; In use, the load prediction value is combined with the real-time environment and equipment information to construct a unified objective function for multi-objective requirements, avoiding the limitations of traditional single energy consumption or single temperature control, and providing well-defined decision variables and objective metrics for subsequent control algorithms; Step 302, Algorithm Design and Real-Time Optimization Define the state vector composed of the predicted value and the current measured environment and equipment parameters , including:

[0034] Define the action vector , corresponding to ; Define the reward function to be consistent with or numerically transformed from the objective function :

[0035] Drive the deep reinforcement learning policy network to continuously optimize the action vector . During the training process, continuously iterate and update the policy network parameters , so that when facing different external environments and load prediction values , it can select an approximately optimal action vector ; In use, through deep reinforcement learning, dynamic scheduling can be achieved under multi-dimensional objectives and coupling constraints, breaking through the complexity limitations of traditional single-loop or P / PI control. Combining load prediction, the upcoming load changes can be incorporated into the decision-making scope, and pre-adjustment of water temperature or valve opening can be carried out in advance, significantly improving the forward-looking and flexibility of control; Utilize the environment-state-action-reward system of deep reinforcement learning to interact with the load prediction model, and introduce forward-looking load information during action decision-making to form an interactive prediction-control framework.

[0036] Step 303, Command Distribution and System-Level Linkage Execution After completing the decision-making of the real-time optimization algorithm or reinforcement learning strategy, the obtained action vector needs to be sent to each sub-device of the air-conditioning system. The action vector , that is, the control instruction, includes: Setting the target outlet water temperature of the chilled water unit , adjusting the frequency of the pump inverter , controlling the opening degree of the chilled water valve ; starting and stopping specific chilled water units or air-conditioning terminal fans ; Through the system master control module or edge device, monitor the execution response time of each device and record its actual operating status for subsequent evaluation of execution deviation, equipment health status, etc.; After actual execution, the updated equipment power , valve opening degree and other equipment operation data can be periodically compared with the predicted load value . If the deviation between the two is greater than the expected value, possible modeling errors or equipment abnormalities can be determined, triggering the online learning mechanism for model adaptation. That is, re-execute the content in step two: When in use, it ensures a fast closed-loop from policy decision-making to on-site execution of the central air-conditioning system, improving the real-time performance of overall control. When there are multiple chilled water units or parallel devices, it can dynamically balance the load distribution between the units, extend the service life of the equipment and reduce the maintenance cost.

[0037] Step Four. When each sub-device completes the execution according to the action vector and sends back the feedback containing indoor environment, equipment power and real-time load data, continuously monitor and evaluate the prediction error and control effect with the help of the operation performance function , and update the parameters of the deep learning model and control algorithm online when the deviation exceeds the threshold; The content of the above-mentioned step four includes: Step 401. Real-time operation effect evaluation and deviation detection Continuously collect the real-time operation status data of the current central air-conditioning, including actual load , indoor and outdoor temperature, equipment power, personnel flow and other information; compare the real-time operation status data with the predicted value , action vector , evaluate the actual execution effect of the control strategy in terms of energy consumption cost and comfort, introduce a custom operation performance function to measure the operation deviation at the current moment, and achieve a more comprehensive evaluation in combination with historical data; Assume that at time it is necessary to evaluate the past time length Define the operation performance function for the impact of as follows:

[0038] In the formula: is the length of the evaluation time window, and the system deviation and control status are cumulatively analyzed within ; is the actual load at time ; is the load prediction value of the deep learning load prediction model at time , and the model parameters are denoted as ; is the amplification coefficient of the power deviation term, used to adjust the overall weight of ; , usually determined according to the sensitivity to the deviation scale; is the power term, used for non-linearly characterizing the deviation , and usually takes ; is the coefficient of the deviation change rate term, used to measure the importance of , and the available range is: ; is the instantaneous change rate of the difference between the actual load and the predicted load; is the penalty or incentive function for the action vector , which is evaluated according to the action intensity, frequency, number of equipment startups and stops, etc.

[0039] When in use, through the continuous calculation of the operation performance function , persistent and high-frequency deviations can be captured, and it can be judged in time whether the current control strategy deviates from the expectation. The sliding window integration form can smooth the instantaneous fluctuations and pay more attention to whether the deviation accumulates to an unacceptable level within a period of time, which is conducive to catching the abnormal signs of equipment or environment. By using the operation performance function to measure the coupling degree between the control deviation and the equipment action, the system can objectively identify when to start the next step of rapid iterative optimization.

[0040] Step 402, Triggering of the rapid iterative optimization process and coordinated update of parameters When it is detected that the operation performance function continuously exceeds the preset threshold within a period of time, it indicates that the deviation of the current load prediction model or control strategy has significantly affected the system energy efficiency or comfort, and it is necessary to trigger the rapid iterative optimization process, including: Retraining of the load prediction model, where: based on the newly collected actual load and the feature vectors at the corresponding moments , perform short-cycle or mini-batch updates to correct the prediction deviation control strategy parameters. Correction: Parameter adjustment of the control strategy, where the objective function or the parameters of the reinforcement learning policy network are iterated in small steps to improve the matching degree for the current working conditions; When in use, the online fast iterative update can make the system continuously converge and adapt to new working conditions during daily operation, without waiting for large-scale offline retraining or manual intervention to achieve synchronous adjustment of multi-step parameters (load prediction model and control strategy or other optimization parameters), avoiding new mismatches in other links caused by only changing a single link; Using the operation performance function as a trigger condition, so that the fast iteration does not only depend on single-point error or instantaneous fluctuation; at the same time, update the load prediction model and the control strategy, creating a collaborative optimization mechanism for cross-step parameters, avoiding the disconnection between the prediction-control links.

[0041] Step 403, Knowledge base accumulation and long-term evolution After each trigger of fast iteration and completion of model and control strategy updates, store the newly learned parameters or control strategy and their corresponding operating environment (such as load characteristics, building conditions, weather conditions, etc.) information into the knowledge base; when subsequent consistent or similar scenarios appear, the similar environment or load conditions in the historical archives can be quickly matched first, and relevant experience and parameter sets can be selected as initial values to improve the iteration update efficiency and reduce the training time; among them, specifically refers to the network parameters used to generate control actions in the deep reinforcement learning module; refers to the set of corrected parameters obtained by updating the original model parameters during the online learning and fast iterative optimization process by comparing the actual operation feedback and prediction results; When in use, the learning results for special scenarios or rare load patterns will not be lost, and the overall operation adaptability level will be improved in the long term. When maintenance personnel need to trace the effect or abnormal cause of a certain control strategy, they can also query the optimal parameters and operating environment during the corresponding period from the knowledge base, which is conducive to further analysis and improvement.

[0042] Step Five. After accumulating the equipment operation trajectories and abnormal records and detecting signs of fault symptoms or performance decay, through the fault symptom function Together with the remote operation and maintenance module, conduct online diagnosis and early warning for core equipment such as chilled water units and pumps, store and synchronize the diagnosis results and maintenance suggestions to the knowledge base to ensure the coordinated promotion of equipment health status and energy-saving strategies; The fifth step includes the following contents: Step 501, Online fault detection and health status analysis Fuse the collected multi-source equipment status data (including pump speed, chilled water unit pressure, valve position, etc.) with the action vector and the energy consumption feedback information to form a real-time status vector for each key equipment; Construct a fault symptom function , within the time window to measure the difference between the current equipment status and the reference health baseline, specifically as follows:

[0043] In the formula: is the multi-dimensional status vector of the rd device at time ; is the health baseline vector of the th device, which can be obtained by statistical, clustering and other methods from historical normal operating condition data; represents the norm of the equipment real-time status and the health baseline (the norm can be selected); is the power coefficient, used for non-linearly amplifying the deviation amplitude, generally taking ; is the rolling time window size of fault detection; is the order of the fractional integral, taking , used to control the weighting degree of data near the current time t; is the Gamma function, related to the definition of fractional integral; is the scaling coefficient, , used to adjust the numerical weight of the fractional integral result; If the fault symptom function continuously exceeds the set fault threshold within a certain period of time, or shows high-frequency fluctuations, that is, trigger a fault warning and report the data and analysis results to the remote operation and maintenance platform; When in use, by performing power integration on the current equipment status and the health baseline within a period of time, it can capture fault symptoms in the early stage, taking into account both continuous micro-deviations and sudden large deviations; echoing with the operation performance function , jointly forming a dual monitoring system for energy consumption deviation and equipment failure. The fault symptom function It can intuitively reflect the health status of each device and provide a decision-making basis for remote operation and maintenance and intervention. By synchronously utilizing four types of information sources, namely environment, control, prediction, and operation status, it overcomes the limitations that may be brought about by solely relying on internal sensors of the device or solely on energy consumption monitoring.

[0044] Step 502, Remote operation and maintenance and fault response strategy When the fault symptom function exceeds the fault threshold and lasts for a certain period of time, a fault warning event is automatically generated and uploaded to the operation and maintenance platform together with data such as the device status vector and the fault symptom trend chart; Based on the knowledge base (including historical maintenance records, similar fault cases, etc.), the operation and maintenance platform provides fault handling suggestions, including maintenance items, required spare parts, or temporary scheduling strategies; To shorten the response cycle and reduce unnecessary downtime losses, temporary control strategy optimization can be carried out before the intervention of operation and maintenance personnel: fine-tune the action vector to give priority to ensuring that the device does not operate under extreme working conditions and avoid the expansion of faults. Only when online fault detection clearly identifies major risks does it enter the remote operation and maintenance and fault response process to reduce the impact of false alarms or frequent interventions on the normal regulation of the system. After a fault event occurs, the fault symptom records and treatment plans will be further supplemented to the knowledge base to provide a quick reference for similar situations that may occur next time. If the faulty device still has partial load capacity, the overall operation of the central air-conditioning system can be maintained by appropriately reducing the power or adjusting the start-stop sequence, thereby reducing the risk of further damage to the device.

[0045] It can seamlessly connect fault diagnosis and remote operation and maintenance, realize the mode of side monitoring, side control, and side maintenance of potential problems, greatly reduce the impact of faults on the system energy efficiency and stability. Through the real-time uploaded data, visual presentation and strategy formulation can be carried out in the cloud. Operation and maintenance personnel can make accurate decisions based on richer data information, avoiding blind disassembly and repeated inspections.

[0046] Those of ordinary skill in the art can realize that, combining the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein, can be implemented by electronic hardware, or by a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of this application.

[0047] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0048] In several embodiments provided in the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only for some logical function divisions. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces. The indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.

[0049] The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0050] As described above, it is only the specific implementation manners of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed in the present application and should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An adaptive control method for central air-conditioning load based on artificial intelligence, characterized in that: include, After each sensor completes a single acquisition, the collected original multi-source data is preprocessed and a feature vector is output; A deep learning model is used to establish load forecasts for multi-dimensional information, and the model parameters are continuously iterated through an adaptive update mechanism to produce real-time forecast results that can dynamically track changes in indoor and outdoor environments. New operation feedback data is collected in real time and compared with the current model forecast results. If it is detected that the loss function of the load forecast model exceeds the preset threshold, the online learning mechanism is started, and the new operation feedback data and its corresponding actual load are included in a small-scale incremental training set. Rapid iteration is performed according to the loss function to update the model parameters. When the prediction results are compared with the actual environmental conditions and form the input conditions, deep reinforcement learning is used to adjust key operating parameters, comprehensively balance energy consumption and comfort, and form intelligent linkage instructions that can proactively adjust the start and stop sequence of each sub-device; When the control command is executed and feedback data is obtained, the forecast error and control effect are continuously monitored in combination with the operation performance function. If it is detected that the threshold is exceeded, the load forecast model and control strategy are modified synchronously. After accumulating real-time and historical data during long-term operation, the fault symptom function and knowledge base linkage mechanism are used to perform online diagnosis of key equipment, and the diagnosis results and maintenance suggestions are stored and synchronized to the knowledge base.

2. The central air conditioning load adaptive control method according to claim 1, characterized in that: Deploy distributed sensors inside and outside the building and send the collected raw data to edge computing nodes; The original data is preliminarily cleaned. If the credibility of the original data is lower than the preset credibility threshold, the edge computing node will mark it as suspicious data.

3. The central air conditioning load adaptive control method according to claim 2, characterized in that: The cleaned and anomaly-detected data are normalized to convert the measured values ​​in the original data into dimensionless feature values; feature weights are assigned to different sensor features and the feature values, their corresponding credibility and weights are summarized into feature vectors.

4. The central air conditioning load adaptive control method according to claim 3, characterized in that: The feature vector is used as the model input. After initializing the deep learning network structure, the weight parameters are initialized randomly or based on prior knowledge. A custom loss function for fusion weight and activation operation is introduced, and the loss function is combined with the optimization algorithm to iteratively update the network parameters until the stopping condition is met.

5. The central air conditioning load adaptive control method according to claim 4, characterized in that: Determine the control objectives of central air conditioning based on load forecast values ​​and the current environmental status data of the building, including: minimizing energy consumption, maintaining indoor temperature and humidity within the set range, and avoiding frequent start-stop or overload operation; Decision variables are established to describe the operating status of each sub-equipment of the central air-conditioning system, including chilled water supply temperature, water pump operating frequency, cold water valve opening, and equipment start and stop instructions. An objective function with adjustable weights is introduced and the objective function is minimized.

6. The central air conditioning load adaptive control method according to claim 5, characterized in that: Define the state vector consisting of the predicted value and the current measured environment and equipment parameters, corresponding to the decision variable action vector; The reward function is defined to be consistent with the objective function or to be numerically transformed, and the reinforcement learning policy network is driven to continuously tune the action vector. The policy network parameters are iteratively updated during the training process, so that it can select the approximately optimal action vector when facing different external environments and load forecast values.

7. The central air conditioning load adaptive control method according to claim 6, characterized in that: The obtained action vector is sent to each sub-device, including: setting the target outlet water temperature of the chilled water unit, adjusting the frequency of the water pump inverter, controlling the opening of the cold water valve, starting and stopping a specific chilled water unit or air conditioner terminal fan; After actual execution, the updated equipment operation data is regularly compared with the predicted load value. If the deviation between the two is greater than expected, it is determined that there is a possible modeling error or equipment abnormality, and the online learning mechanism is triggered to perform model adaptation.

8. The central air conditioning load adaptive control method according to claim 7, characterized in that: Continuously collect the real-time operation status data of the current central air conditioner, and compare the real-time operation status data with the load forecast value and action vector; A custom operation performance function is introduced to measure the operation deviation at the current moment. When it is detected that the operation performance function exceeds the preset threshold for a period of time, a rapid iterative optimization process is triggered, including: Based on the recently collected actual load and the feature vector at the corresponding moment, short-cycle or micro-batch updates are performed to correct the prediction deviation control strategy parameters, and small-step iterations are performed on the objective function or reinforcement learning strategy network parameters.

9. The central air conditioning load adaptive control method according to claim 8, characterized in that: After each rapid iteration is triggered and the model and control strategy are updated, the newly learned parameters or action vectors and their corresponding operating environment information are stored in the knowledge base; When the same or similar scenarios occur in the future, you can quickly match similar environments or load conditions in the historical archives and select relevant experience and parameter sets as initial values.

10. The central air conditioning load adaptive control method according to claim 9, characterized in that: The collected multi-source equipment status data is integrated with the action vector and energy consumption feedback information to form a real-time state vector for each key equipment; a fault symptom function is constructed to measure the difference between the current equipment status and the reference health baseline within a time window; If the fault symptom function continues to exceed the set fault threshold for a period of time, or high-frequency fluctuations occur, a fault warning is triggered and the data and analysis results are reported to the remote operation and maintenance platform.

11. The central air conditioning load adaptive control method according to claim 10, characterized in that: When the fault symptom function exceeds the fault threshold and lasts for the target duration, a fault warning event is automatically generated and uploaded to the operation and maintenance platform along with the device state vector and fault symptom trend chart; The operation and maintenance platform provides troubleshooting suggestions based on the knowledge base, including maintenance items, required parts or temporary scheduling strategies; Perform temporary control strategy tuning and fine-tune the action vector before operation and maintenance personnel intervene.

Images