Artificial intelligence-based constant temperature and humidity air conditioning energy-saving regulation method and related equipment

Abstract

The application discloses an energy-saving regulation and control method for constant temperature and humidity air conditioners based on artificial intelligence and related equipment, and relates to the technical field of artificial intelligence. The method comprises the following steps: collecting environment data of a constant temperature and humidity space and operation state data of an air conditioning device; inputting the environment data and the operation state data into a temperature and humidity coupling prediction model and an energy consumption prediction model to output a temperature dynamic prediction trend, a humidity dynamic prediction trend and an energy consumption change trend in a future control period; calculating temperature and humidity prediction uncertainty, and tightening preset control boundary parameters when the temperature and humidity prediction uncertainty is higher than a preset threshold; constructing a multi-objective function under the constraint of the control boundary parameters; generating a regulation and control strategy by using a reinforcement learning model of a fusion model prediction control based on the multi-objective function and the energy consumption change trend; and distributing the regulation and control strategy to corresponding execution components for hierarchical control according to the response time of the multiple execution components of the air conditioning device.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an energy-saving control method and related equipment for constant temperature and humidity air conditioning based on artificial intelligence. Background Technology

[0002] With the development of industries such as precision manufacturing, biomedicine, and high-tech laboratories, constant temperature and humidity air conditioning systems play a crucial role in maintaining the stability of spatial environments. However, in order to simultaneously meet stringent temperature and humidity standards, these systems typically require high-load operation, resulting in persistently high energy consumption. Improving energy efficiency while maintaining environmental precision remains a continuous challenge in this field.

[0003] Currently, the control of constant temperature and humidity air conditioners largely relies on traditional proportional-integral-derivative (PID) algorithms, or decouples temperature and humidity as independent variables. This approach often leads to the phenomenon of mutual cancellation of heating and cooling, such as frequent activation of reheat devices due to excessive dehumidification during cooling, resulting in energy waste. Furthermore, some data-driven predictive control schemes in existing technologies are prone to prediction failure when faced with strong external disturbances or under data-sparse operating conditions, leading to temperature and humidity exceeding limits and compromising environmental safety and stability. Simultaneously, at the equipment scheduling level, existing automation systems typically tend to use a single control cycle to synchronously schedule all actuators. This can easily cause frequent mechanical movements and additional component wear when dealing with physical components with varying response speeds.

[0004] Therefore, there is an urgent need for an energy-saving control technology that can dynamically adapt to complex working conditions while taking into account the response characteristics of the equipment. Summary of the Invention

[0005] This invention provides an energy-saving control method and related equipment for constant temperature and humidity air conditioning based on artificial intelligence, which solves the problems of high energy consumption, severe cooling and heating offsetting, and easy occurrence of temperature and humidity exceeding limits in the existing technology.

[0006] The first aspect of this invention proposes an energy-saving control method for constant temperature and humidity air conditioning based on artificial intelligence, comprising: Collect environmental data of the constant temperature and humidity space and the operating status data of the air conditioning equipment; Environmental data and operational status data are input into a pre-trained temperature and humidity coupled prediction model and an energy consumption prediction model, which respectively output the dynamic prediction trends of temperature, humidity and energy consumption in the future control cycle. Based on the dynamic prediction trends of temperature and humidity, the uncertainty of temperature and humidity prediction is calculated, and when the uncertainty of temperature and humidity prediction exceeds the preset uncertainty threshold, the preset control boundary parameters are tightened. Construct a multi-objective function under the constraints of the control boundary parameters; Based on multiple objective functions and energy consumption change trends, a reinforcement learning model for predictive control using a fusion model is adopted to generate regulation strategies. Based on the response time of the multiple actuators contained in the air conditioning equipment, the control strategy is allocated to the corresponding actuators for hierarchical control.

[0007] In one optional implementation, environmental data of the constant temperature and humidity space and operating status data of the air conditioning equipment are collected, including: The constant temperature and humidity space is divided into multiple independent control zones; The indoor temperature, outdoor temperature, relative humidity, dew point temperature, supply and return air status, space load, personnel heat generation, and process equipment heat generation of each control area are collected in real time as environmental data. Simultaneously collect the operating power, compressor frequency, fan frequency, chilled water valve opening, humidification status, dehumidification status, and reheat status of each of the multiple actuators as operating status data; Environmental data and operational status data are uploaded to a cloud server, and a remote control strategy is generated through a reinforcement learning model that combines temperature and humidity coupling prediction model, energy consumption prediction model, and fusion model prediction and control within the cloud server.

[0008] In one optional implementation, environmental data and operational status data are input into a pre-trained temperature and humidity coupled prediction model and an energy consumption prediction model, which respectively output the dynamic prediction trends of temperature, humidity, and energy consumption within future control cycles, including: Environmental data and operational status data are cleaned and their features are standardized to construct a fused feature vector that includes environmental status features, equipment action features, and historical response features. The fused feature vector is input into the temperature and humidity coupled prediction model. The graph neural network in the temperature and humidity coupled prediction model is used to extract the spatial thermodynamic interaction correlation features in the constant temperature and humidity space, and the long short-term memory network is used to extract the dynamic temporal dependence features of temperature and humidity changes over time. The spatial thermal interaction correlation features and dynamic temporal dependence features are weighted and fused through the attention mechanism, and the nonlinear coupling dynamic response relationship between temperature and humidity is established through the latent variable sharing layer. The output includes the temperature dynamic prediction trend containing the temperature mean and temperature variance, and the humidity dynamic prediction trend containing the humidity mean and humidity variance. The fused feature vector is input into the energy consumption prediction model, and the correlation between device state changes and energy consumption is mapped through a multilayer perceptron network, outputting an energy consumption change trend that includes the expected power consumption.

[0009] In one optional implementation, the temperature and humidity prediction uncertainty is calculated based on the dynamic prediction trends of temperature and humidity. When the temperature and humidity prediction uncertainty exceeds a preset uncertainty threshold, preset control boundary parameters are tightened, including: From the dynamic prediction trends of temperature and humidity, a prediction variance matrix containing the prediction variance is extracted, and the trace of the prediction variance matrix is ​​used as the uncertainty of temperature and humidity prediction to characterize the confidence of the model prediction. Determine whether the uncertainty of temperature and humidity forecasts exceeds the uncertainty threshold; When the uncertainty of temperature and humidity prediction exceeds the uncertainty threshold, the current system state is determined to be in a dynamic edge condition with sparse data, triggering the over-limit risk penalty mechanism. The boundary contraction coefficient is calculated based on the ratio of the difference between the uncertainty of temperature and humidity prediction and the uncertainty threshold. By using the boundary contraction coefficient, the upper and lower boundaries of the preset temperature tolerance range and humidity tolerance range in the control boundary parameters are compressed proportionally to obtain the tightened control boundary parameters. When the uncertainty of temperature and humidity prediction is not higher than the uncertainty threshold, the initial temperature tolerance range and humidity tolerance range are maintained as control boundary parameters.

[0010] In one optional implementation, a multi-objective function is constructed, and a reinforcement learning model for predictive control based on the multi-objective function and energy consumption change trends is used to generate a regulation strategy, including: A multi-objective function is constructed with the objectives of minimizing temperature deviation, humidity deviation, energy consumption, and the number of equipment actions. The dynamic prediction trends of temperature, humidity, energy consumption, and control boundary parameters are used as the environmental state space and input into the reinforcement learning model of fusion model predictive control. Starting from the current environmental state space, the model prediction path integral algorithm is used to introduce Gaussian noise for random sampling in the preset continuous action space to generate a preset number of candidate action sequences. The comprehensive cost of each candidate action sequence in the future control cycle is calculated using a multi-objective function. The comprehensive cost is obtained by weighted summation of temperature deviation cost, humidity deviation cost, energy consumption prediction cost, and equipment operation frequency penalty cost. When the temperature and humidity values ​​predicted by the evaluated candidate action sequence exceed the control boundary parameters, the corresponding temperature deviation cost and humidity deviation cost are given a step penalty weight. Based on the comprehensive cost value of all candidate action sequences obtained from the calculation, calculate the information theory probability weight of each candidate action sequence. The probability weights of information theory are used to perform a weighted average of all candidate action sequences and then merge them to generate the globally optimal action sequence. Extract the set of optimal actions corresponding to the first control cycle in the globally optimal action sequence, and generate a control strategy containing multiple control actions.

[0011] In one optional implementation, the control strategy is allocated to the corresponding actuators for hierarchical control based on the response time of the multiple actuators included in the air conditioning equipment, including: Based on the mechanical inertia, thermodynamic response delay time, and control constraint complexity of each of the multiple actuators, the multiple actuators are divided into a slow response layer, a medium response layer, and a fast response layer. The slow response layer is configured to include compressor start / stop control components for the basic cold and heat source and chilled water main valve regulating components, and the control frequency of the slow response layer is set to the first cycle; The medium-speed response layer is configured to include a fan frequency regulation component and a reheat compensation device for coarse adjustment of sensible heat load and latent heat load, and the control frequency of the medium-speed response layer is set to a second cycle that is less than the first cycle. The fast response layer is configured to include a humidification device and a dehumidification device for precise temperature and humidity adjustment, and the control frequency of the fast response layer is set to a third cycle that is less than the second cycle. By using a hierarchical reinforcement learning framework, the control strategy is decomposed into a first control instruction decided by a high-level agent, a second control instruction decided by a mid-level agent, and a third control instruction decided by a low-level agent according to the time scale. The first control command is sent to the slow response layer, the second control command is sent to the medium response layer, and the third control command is sent to the fast response layer, so that the execution components with different response speeds can perform actions asynchronously within their respective cycles, thereby completing the hierarchical control.

[0012] In one alternative implementation, the method further includes: During the implementation of hierarchical control, the current actual temperature and humidity data of the constant temperature and humidity space are obtained in real time through continuous collection of environmental data. Compare the current actual temperature and humidity data with the pre-set hard safety protection threshold range for the constant temperature and humidity space; If the current actual temperature data or the current actual humidity data exceeds the hard safety protection threshold, it is determined that the current system has encountered a sudden strong disturbance or that the control strategy has a risk of exceeding the limit, and the control strategy generated by the reinforcement learning model of the fusion model predictive control is immediately suspended. The control mode is automatically switched to a conservative control strategy based on the proportional-integral-derivative algorithm to prioritize restoring the temperature and humidity of the constant temperature and humidity space to within the safe baseline. Collect abnormal state transition data at the moment when the hard safety protection threshold is triggered, and store the abnormal state transition data in a preset experience playback pool for offline retraining of the reinforcement learning model of temperature and humidity coupled prediction model and fusion model prediction control.

[0013] A second aspect of this invention provides an artificial intelligence-based energy-saving control device for constant temperature and humidity air conditioning, comprising: The data acquisition module is used to collect environmental data of the constant temperature and humidity space and the operating status data of the air conditioning equipment; The dynamic prediction module is used to input environmental data and operating status data into the pre-trained temperature and humidity coupled prediction model and energy consumption prediction model, and output the dynamic prediction trend of temperature, dynamic prediction trend of humidity and energy consumption change trend in the future control cycle, respectively. The boundary constraint module is used to calculate the temperature and humidity prediction uncertainty based on the temperature dynamic prediction trend and the humidity dynamic prediction trend, and tighten the preset control boundary parameters when the temperature and humidity prediction uncertainty is higher than the preset uncertainty threshold. The objective building module is used to construct multi-objective functions under constraints of controlling boundary parameters; The strategy generation module is used to generate regulation strategies based on a reinforcement learning model that employs fusion model predictive control, using a multi-objective function and energy consumption change trends. The hierarchical control module is used to allocate control strategies to the corresponding actuators for hierarchical control based on the response time of the multiple actuators contained in the air conditioning equipment.

[0014] A third aspect of the present invention provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; The memory stores a computer program that can be executed by at least one processor. When the computer program is executed by at least one processor, it implements the energy-saving control method for constant temperature and humidity air conditioning based on artificial intelligence as proposed in the foregoing embodiments.

[0015] The fourth aspect of this invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the artificial intelligence-based energy-saving control method for constant temperature and humidity air conditioning as proposed in the foregoing embodiments.

[0016] Beneficial Effects: This invention provides an AI-based energy-saving control method and related equipment for constant temperature and humidity air conditioning. By establishing a coupled temperature and humidity prediction and energy consumption prediction mechanism, and combining model predictive control and reinforcement learning, it achieves multi-objective optimization under control boundary constraints, overcoming the problems of severe heat and cold offsetting and rigid control modes in existing technologies. Simultaneously, an uncertainty assessment mechanism for temperature and humidity prediction is introduced. When facing data-sparse operating conditions, tightening the control boundary effectively reduces the risk of environmental runaway caused by prediction deviations, improving the system's over-limit protection capability and control safety. Furthermore, hierarchical control is implemented based on the response time differences of various components in the air conditioning equipment, allowing components with different response speeds to operate asynchronously within their respective suitable cycles. This reduces redundant mechanical movements and ineffective energy consumption, helping the system maintain a stable and efficient control state under complex operating environments. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention; Figure 2 This is a flowchart illustrating the steps of an artificial intelligence-based energy-saving control method for constant temperature and humidity air conditioning, as provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the functional modules of an artificial intelligence-based constant temperature and humidity air conditioning energy-saving control device provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0019] The present invention will be further described below with reference to the accompanying drawings.

[0020] Reference Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention.

[0021] like Figure 1As shown, the electronic device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless-fidelity (Wi-Fi) interface. The memory 1005 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0022] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0023] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a data storage module, a network communication module, a user interface module, and computer programs.

[0024] exist Figure 1 In the electronic device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the electronic device of the present invention can be set in the electronic device. The electronic device calls the AI-based constant temperature and humidity air conditioning energy-saving control device stored in the memory 1005 through the processor 1001 and executes the AI-based constant temperature and humidity air conditioning energy-saving control method provided in the embodiment of the present invention.

[0025] Combined with appendix Figure 2 As shown, the first aspect of this invention proposes an energy-saving control method for constant temperature and humidity air conditioning based on artificial intelligence. This method aims to solve the multi-objective optimization problem between high-precision environmental maintenance and deep energy saving in complex physical spaces, and specifically includes the following steps: S101 collects environmental data of the constant temperature and humidity space and the operating status data of the air conditioning equipment; In complex constant temperature and humidity application scenarios (such as semiconductor lithography workshops, high-level biosafety laboratories, or precision manufacturing workshops), even minor disturbances in environmental parameters can lead to a decrease in process yield or distortion of experimental data. Therefore, multi-dimensional ubiquitous sensing is the physical basis for achieving intelligent control. In one optional implementation method, environmental data from the constant temperature and humidity space and the operating status data of the air conditioning equipment are collected, including: S1011. Divide the constant temperature and humidity space into multiple independent control zones; S1012. Collect in real time the indoor temperature, outdoor temperature, relative humidity, dew point temperature, supply and return air status, space load, personnel heat generation, and process equipment heat generation of each control area as environmental data. S1013. Simultaneously collect the operating power, compressor frequency, fan frequency, chilled water valve opening, humidification status, dehumidification status, and reheat status of each of the multiple actuators as operating status data; S1014. Upload environmental data and operational status data to the cloud server to remotely generate control strategies through the reinforcement learning model of temperature and humidity coupling prediction model, energy consumption prediction model and fusion model prediction control in the cloud server.

[0026] In step S1011 above, the specific implementation process of dividing the constant temperature and humidity space into multiple independent control areas includes: acquiring the building information model (BIM) of the constant temperature and humidity space, extracting the physical partition materials, thermal resistance parameters, and physical coordinates of the air supply and return vents within the space. Based on the extracted parameters, a preliminary three-dimensional flow field and temperature field simulation is performed using computational fluid dynamics (CFD) simulation technology. Based on the connectivity of the flow field distribution and the degree of temperature gradient aggregation, a large physical constant temperature and humidity space is logically divided into multiple control areas with significant thermodynamic independence or weak coupling. Each control area is configured as an independent state perception and environmental feedback node to facilitate subsequent refined spatial modeling.

[0027] In step S1012 above, to obtain high-fidelity environmental data, an Internet of Things (IoT) sensing array containing various types of sensors is deployed in each independently defined control area. Indoor temperature and relative humidity are acquired at a high-frequency sampling rate using high-precision platinum resistance temperature probes and capacitive humidity probes placed on workbenches, ceilings, and key return air paths. Simultaneously, outdoor temperature feedback from an outdoor weather station is acquired, and the current dew point temperature is calculated using the water vapor partial pressure principle based on the real-time indoor temperature and relative humidity for subsequent safety checks to prevent condensation. The supply and return air status is comprehensively quantified using hot-wire anemometers and micro-differential pressure sensors installed in the air conditioning ducts. For dynamic changes in space load, the number of people in the current area is extracted through the access control system's people flow statistics interface and multiplied by the average heat dissipation base per person to obtain the heat generated by each person; the real-time power consumption of each process device in the area is read through the industrial communication bus, and the heat generated by the process device is calculated based on the law of conservation of energy. All the values ​​obtained through physical sensor acquisition and logical calculation constitute the environmental data required by this method.

[0028] In step S1013 above, the executing entity polls the underlying programmable logic controller of the air conditioning equipment via an industrial communication protocol (such as Modbus or the Programmable Instrument Standard Command Protocol) to read the operating status data of multiple executing components. This data precisely reflects the current physical working state of the air conditioning system, specifically covering: the current operating frequency of the variable frequency compressor, which characterizes the basic cooling and heating output capacity; the current operating frequency of the variable frequency fan, which characterizes the air circulation and heat exchange rate; the opening degree of the chilled water valve under proportional-integral control, which characterizes the cooling capacity delivery efficiency; the humidification status of the humidification device (such as an electrode humidifier or ultrasonic humidifier), which characterizes the latent heat treatment capacity; the dehumidification status, which characterizes the air dehydration capacity; and the reheat status, which characterizes the reheat device (such as an electric heater or a hot gas bypass circuit), which characterizes the reheating device to prevent overcooling.

[0029] In step S1014 above, given that local control nodes typically face computing power bottlenecks and cannot efficiently perform inference calculations for deep neural networks, an edge gateway with data encryption and breakpoint resume capabilities packages the collected environmental and operational status data and securely uploads it to a cloud server with Graphics Processing Unit (GPU) cluster computing power via a dedicated fiber optic line or 5G mobile communication technology. Leveraging its computing power advantage, the cloud server hosts and runs complex nonlinear mathematical models, thereby enabling remote high-order control through cloud-edge collaboration.

[0030] S102, input environmental data and operating status data into the pre-trained temperature and humidity coupled prediction model and energy consumption prediction model, and output the dynamic prediction trend of temperature, dynamic prediction trend of humidity and energy consumption change trend in the future control cycle respectively. In heating, ventilation, and air conditioning (HVAC) thermodynamic systems, temperature and humidity are not isolated variables. For example, while a surface cooler lowers the air temperature (handles sensible heat), it often condenses water vapor in the air, thereby lowering the absolute humidity (handling latent heat); conversely, the humidification process is often accompanied by the absorption or release of heat. Therefore, both must be considered as a coupled system for joint prediction. In one optional implementation, step S102 specifically includes: S1021. Perform data cleaning and feature standardization on environmental data and operational status data to construct a fusion feature vector that includes environmental status features, equipment action features, and historical response features; S1022. Input the fused feature vector into the temperature and humidity coupled prediction model, use the graph neural network in the temperature and humidity coupled prediction model to extract the spatial thermodynamic interaction correlation features in the constant temperature and humidity space, and use the long short-term memory network to extract the dynamic temporal dependence features of temperature and humidity changing over time. S1023. The spatial thermal interaction correlation features and dynamic temporal dependence features are weighted and fused through the attention mechanism, and the nonlinear coupling dynamic response relationship between temperature and humidity is established through the latent variable sharing layer. The output includes the temperature dynamic prediction trend containing the temperature mean and temperature variance, and the humidity dynamic prediction trend containing the humidity mean and humidity variance. S1024. Input the fused feature vector into the energy consumption prediction model, and map the correlation between device state changes and energy consumption through a multilayer perceptron network to output the energy consumption change trend containing the expected power consumption.

[0031] In step S1021 above, the data undergoes rigorous preprocessing before entering the model. The execution unit first uses a moving median filter algorithm or a density-based spatial clustering algorithm to remove outliers caused by high-frequency noise or electromagnetic interference from sensors; for missing data, Lagrange interpolation is used to complete it. Subsequently, a min-max normalization technique is used to map physical quantities with different dimensions (such as temperature in degrees Celsius, frequency in Hertz) to a dimensionless interval of zero to one, eliminating the negative impact of feature scale on the gradient descent of the neural network. The normalized current environmental data is defined as environmental state features, the current equipment operating state data is defined as equipment action features, and historical state sequences within a preset time window are extracted as historical response features. These three types of features are concatenated along the tensor dimension to construct a high-dimensional fusion feature vector for the input model.

[0032] In step S1022 above, a Graph Neural Network (GNN) is used to capture the topological influence of the spatial dimension. In the structural definition of the GNN, each control region divided in step S1011 is abstracted as a graph node. If there is a physical adjacency or air convection path between two control regions, a directed edge is established between the two nodes, and the weight of the edge is initialized with the thermal conductivity coefficient and airflow rate between them. Through multiple iterations of node feature aggregation and information transmission, the GNN enables the feature vector of each region to deeply integrate the thermal radiation and cold and hot air diffusion effects of the surrounding areas, thereby extracting the global spatial thermal interaction correlation features. In parallel, given the strong temporal hysteresis effect of the thermodynamic system, a Long Short-Term Memory (LSTM) network is used. The LSTM network uses its internal forget gate, input gate, and output gate structure to control the retention and forgetting of historical information, extracting the dynamic temporal dependence features of heat accumulation and decay from the sequential data of the time window.

[0033] In step S1023 above, an attention mechanism is introduced to align the extracted features. The attention mechanism dynamically assigns attention weights by calculating the correlation scores between spatial and temporal features when predicting the current state, and then adds or concatenates the weighted features. Subsequently, the fused features enter the Shared Latent Variable Layer. This layer, composed of a fully connected neural network, maps the input features to a high-dimensional implicit representation space. Since both the temperature and humidity prediction outputs extract information from this same latent space, this architecture forces the network to learn the intrinsic coupling constraints between temperature and humidity in physical processes such as evaporation and condensation during training, avoiding the physical paradoxes caused by independent predictions. Finally, to quantify the prediction confidence, the output adopts a probability distribution prediction method. For each future control step, it outputs the mean temperature and mean humidity representing the expected trend, as well as the variance of temperature and variance of humidity representing the predicted dispersion range, collectively constituting the dynamic prediction trends of temperature and humidity.

[0034] In step S1024 above, the independent energy consumption prediction model is mainly composed of a Multi-Layer Perceptron (MLP) network. This network deeply fits complex nonlinear mapping relationships, including the variable frequency power consumption curve of the compressor, the energy consumption characteristics of the fan according to the cubic law of hydrodynamics, and the linear resistive power consumption characteristics of the reheat device, through multi-layer fully connected computation and nonlinear activation function mapping. Finally, it outputs the numerical trajectory of the electrical energy expected to be consumed in the future control cycle, that is, the energy consumption change trend.

[0035] S103, calculate the temperature and humidity prediction uncertainty based on the temperature dynamic prediction trend and the humidity dynamic prediction trend, and tighten the preset control boundary parameters when the temperature and humidity prediction uncertainty is higher than the preset uncertainty threshold. Deep learning models are highly susceptible to making overconfident and erroneous predictions when encountering out-of-distribution data (such as extreme weather events or sudden large-scale heat source access). Blindly accepting these predictions and generating control commands based on them could lead to uncontrolled environmental parameters. In one alternative implementation, security defense steps based on an uncertainty awareness mechanism include: S1031. Extract the prediction variance matrix containing the prediction variance from the dynamic prediction trends of temperature and humidity, and use the trace of the prediction variance matrix as the temperature and humidity prediction uncertainty that characterizes the prediction confidence of the model. S1032. Determine whether the uncertainty of temperature and humidity prediction is higher than the uncertainty threshold. S1033. When the uncertainty of temperature and humidity prediction exceeds the uncertainty threshold, the current system state is determined to be in a dynamic edge condition with sparse data, triggering the limit-breaking risk penalty mechanism. S1034. Calculate the boundary contraction coefficient based on the ratio of the difference between the uncertainty of temperature and humidity prediction and the uncertainty threshold. S1035. Using the boundary contraction coefficient, the upper and lower boundaries of the preset temperature tolerance range and humidity tolerance range in the control boundary parameters are compressed proportionally to obtain the tightened control boundary parameters. S1036. When the uncertainty of temperature and humidity prediction is not higher than the uncertainty threshold, the initial temperature tolerance range and humidity tolerance range shall be maintained as control boundary parameters.

[0036] In step S1031 above, the executing entity uses the variance information output in S1023 to construct a prediction variance matrix with the prediction variances of each predictor variable (such as temperature and humidity in each control area) as its diagonal elements. In matrix analysis theory, the sum of the main diagonal elements of a matrix (i.e., the trace of the matrix) can macroscopically characterize the overall dispersion of the entire multivariate probability distribution. Therefore, by calculating the trace of this prediction variance matrix, it is used as a scalar indicator characterizing the reciprocal of the model's current prediction confidence, i.e., the uncertainty of temperature and humidity prediction. The larger this value, the less confident the model is in predicting future trends.

[0037] In steps S1033 and S1034 above, once it is determined that the scalar index is higher than the preset uncertainty threshold, it indicates that the current state of the controlled environment deviates from the model's historical cognitive domain. At this point, a risk defense mechanism is triggered. To quantify this risk and smooth the transition, the difference between the current temperature and humidity prediction uncertainty and the uncertainty threshold is calculated, and this difference is normalized by dividing by the uncertainty threshold. Subsequently, the normalized value is input into a preset linear decay function or negative exponential mapping function, outputting a constant multiplier between zero and one, defined as the boundary contraction coefficient. The higher the degree of uncertainty, the smaller the output boundary contraction coefficient (closer to zero).

[0038] In step S1035 above, the control boundary parameters were originally set to the physical limits allowed by the constant temperature and humidity process (for example, setting the target temperature to 22 degrees Celsius and the tolerance to ±1 degree Celsius, then the initial upper boundary is 23 degrees Celsius and the lower boundary is 21 degrees Celsius). After obtaining the boundary contraction coefficient, the initial tolerance span is multiplied by the contraction coefficient and reduced, thereby forcing the upper boundary to move downward and the lower boundary to move upward. Through this proportional space compression operation, the system artificially and forcibly defines a "narrow active zone" that is more stringent than the process requirements for subsequent optimization algorithms. This conservative boundary tightening operation helps to reserve sufficient buffer space when the model lacks confidence, preventing the actual physical environment from exceeding the process limits due to prediction overshoot.

[0039] In step S1036 above, if the uncertainty is low, it proves that the model is fully "familiar" with the current working conditions and has high prediction accuracy. In this case, the initial tolerance range is kept relatively wide, and the optimization algorithm is given the maximum degree of freedom to fully tap the energy-saving potential.

[0040] S104, under the constraints of control boundary parameters, a multi-objective function is constructed; the multi-objective function is used in the subsequent S105 to perform comprehensive cost calculation on the candidate action sequence. S105, based on multiple objective functions and energy consumption change trends, adopts a reinforcement learning model of fusion model predictive control to generate regulation strategies; Due to the strong temporal inertia and vast continuous action solution space of HVAC equipment, traditional optimization algorithms often fail to meet real-time control requirements due to excessive computation time or getting trapped in local optima. In one alternative implementation, policy generation is achieved through the following fusion mechanism: Model Predictive Path Integral (MPPI). S1051. Construct a multi-objective function with the objectives of minimizing temperature deviation, minimizing humidity deviation, minimizing energy consumption, and minimizing the number of equipment actions. S1052. The dynamic prediction trends of temperature, humidity, energy consumption, and control boundary parameters are used as the environmental state space and input into the reinforcement learning model of the fusion model predictive control. S1053. Starting from the current environmental state space, the model prediction path integral algorithm is used to introduce Gaussian noise for random sampling in the preset continuous action space to generate a preset number of candidate action sequences. S1054. Calculate the comprehensive cost value of each candidate action sequence in the future control cycle using a multi-objective function. The comprehensive cost value is obtained by weighted summation of temperature deviation cost, humidity deviation cost, energy consumption prediction cost, and equipment action frequency penalty cost. When the temperature and humidity values ​​predicted by the evaluated candidate action sequence exceed the control boundary parameters, the corresponding temperature deviation cost and humidity deviation cost are given a step penalty weight. S1055. Based on the comprehensive cost value of all candidate action sequences obtained from the calculation, calculate the information theory probability weight of each candidate action sequence. S1056. Use information theory probability weights to perform a weighted average of all candidate action sequences and fuse them to generate the globally optimal action sequence. S1057. Extract the set of optimal actions corresponding to the first control cycle in the global optimal action sequence, and generate a control strategy containing multiple control actions.

[0041] In step S1051 above, the constructed multi-objective function aims to balance the contradictory relationships among the four factors. Its mathematical logic is as follows: the absolute values ​​of the deviations between the predicted temperature and the target temperature, the absolute values ​​of the deviations between the predicted humidity and the target humidity, the predicted total energy consumption, and the step difference (characterizing the frequency of action) before and after the change of equipment control commands at each time step are multiplied by their respective preset weighting coefficients and then added together to form a comprehensive evaluation index.

[0042] In step S1053 above, during the optimization process, the MPPI algorithm, starting from the current physical state, injects thousands of independent Gaussian white noises in parallel into a pre-defined baseline action sequence based on the Monte Carlo random sampling concept. Through this random perturbation exploration, hundreds or thousands of different hypothetical control trajectories are instantaneously "diverged" outwards along the future timeline within the algorithm. Each trajectory represents a possible combination of actions that can be executed consecutively, i.e., multiple sets of candidate action sequences.

[0043] In step S1054 above, a pre-trained temperature and humidity coupled prediction model and energy consumption prediction model are used as a virtual forward inference simulator to calculate the future trend of each candidate action sequence. These are then substituted into a multi-objective function for scoring. During this process, a safety threshold of "soft constraints turning into hard penalties" is set: if the forward inference finds that an action sequence will cause the predicted temperature and humidity values ​​at a certain time step to exceed the tightened control boundary parameters in S1035, the system immediately multiplies the temperature deviation cost and humidity deviation cost of that candidate sequence by a very large step constant penalty factor. This operation causes any dangerous sequence that attempts to sacrifice environmental limits for low energy consumption to rapidly increase its overall cost, thus eliminating it from subsequent screening.

[0044] In steps S1055 and S1056 above, according to information theory principles, each calculated comprehensive cost value is negatively charged and then exponentially calculated. This operation ensures that the action sequence with a smaller cost value has a larger result after exponential mapping. Normalizing all exponential calculation results yields the information theory probability weights for each candidate action sequence. Subsequently, these probability weights are used as multipliers to perform a weighted average summation of all candidate action sequences perturbed by Gaussian noise. This process automatically filters out adverse action directions that lead to high cost values, while positively and smoothly superimposing beneficial control actions pointing to low-energy, high-stability directions, thus fusing a globally optimal action sequence in a continuous high-dimensional space without gradient interference.

[0045] In step S1057 above, following the rolling horizon principle of Model Predictive Control (MPC), the executing entity extracts only the action instruction combination corresponding to the "first control cycle" that needs to be executed immediately from this long-term globally optimal action sequence, and packages it to generate a control strategy. As time progresses, when the next control cycle arrives, the entire state evaluation and sampling optimization process will be retried based on the latest feedback environmental data.

[0046] S106, based on the response time of the multiple actuators included in the air conditioning equipment, the control strategy is allocated to the corresponding actuators for hierarchical control.

[0047] If the control strategy is uniformly applied to all components within the unit at the same time, the difference in rotational inertia between the fluid machinery and the phase change time of the heat exchange medium can easily lead to thermal oscillations and mechanical wear. In one alternative implementation, decoupling is achieved through the following steps based on Hierarchical Reinforcement Learning (HRL): S1061. Based on the mechanical inertia, thermodynamic response delay time, and control constraint complexity of each of the multiple actuators, the multiple actuators are divided into a slow response layer, a medium response layer, and a fast response layer. S1062. Configure the slow response layer to include compressor start / stop control components and chilled water main valve regulating components for the supply of basic cold and heat sources, and set the control frequency of the slow response layer to the first cycle; S1063. The medium-speed response layer is configured to include a fan frequency adjustment component and a reheat compensation device for coarse adjustment of sensible heat load and latent heat load, and the control frequency of the medium-speed response layer is set to a second cycle that is less than the first cycle. S1064. Configure the fast response layer to include a humidification device and a dehumidification device for precise temperature and humidity adjustment, and set the control frequency of the fast response layer to a third cycle that is less than the second cycle. S1065. Through a hierarchical reinforcement learning framework, the control strategy is decomposed into a first control instruction decided by a high-level agent, a second control instruction decided by a mid-level agent, and a third control instruction decided by a low-level agent according to the time scale. S1066. Send the first control command to the slow response layer, the second control command to the medium response layer, and the third control command to the fast response layer, so that the execution components with different response speeds can perform actions asynchronously in their respective cycles to complete the hierarchical control.

[0048] In steps S1061 to S1064 above, based on the intrinsic characteristics of fluid mechanics, changes in the compressor and chilled water main valve will cause a drastic redistribution of pressure throughout the pipeline. Frequent operation can easily lead to surge or reduced lifespan. Therefore, it is classified into the slow response layer, with the longest first cycle (e.g., 15 minutes) set to maintain a stable supply of cooling capacity to the base plate. The variable frequency fan can quickly respond to sudden changes in indoor sensible heat load by changing its speed, so it is classified into the medium response layer and matched with a shorter second cycle (e.g., 3 minutes). The start-stop response of the humidification device and specific dehumidification branch circuits has almost no mechanical delay, so it is classified into the fast response layer and matched with an extremely short third cycle (e.g., 30 seconds) for high-frequency peak shaving and valley filling fine-tuning.

[0049] In steps S1065 and S1066 above, leveraging the dimensionality reduction characteristics of the hierarchical framework, the control strategy is logically decomposed into sub-instructions at different time scales. Through a specially designed hardware communication scheduling middleware, the first, second, and third control instructions are pushed into different communication queues according to the clock cycles set for each layer. This operation completely decouples components with different characteristics in the time domain, allowing them to execute actions asynchronously. This ensures both the stability and longevity of the basic equipment and the sensitivity of the system's endpoints, thus achieving hierarchical control that balances equipment safety and environmental accuracy.

[0050] To further enhance the safety redundancy of the control process, the method also includes fallback safety logic and closed-loop evolution steps.

[0051] In one alternative implementation, the method further includes: S1071. During the execution of hierarchical control, the current actual temperature and humidity data of the constant temperature and humidity space are obtained in real time through continuous collection of environmental data. S1072. Compare the current actual temperature data and current actual humidity data with the hard safety protection threshold range that has been set in advance for the constant temperature and humidity space; S1073. If the current actual temperature data or the current actual humidity data exceeds the hard safety protection threshold range, it is determined that the current system has encountered a sudden strong disturbance or the control strategy has a risk of exceeding the limit, and the control strategy generated by the reinforcement learning model of the fusion model predictive control is suspended. S1074. Automatically switch the control mode to a basic conservative control strategy based on proportional-integral-derivative algorithm to prioritize restoring the temperature and humidity of the constant temperature and humidity space to within the safe baseline. S1075. Collect abnormal state transition data at the moment when the hard safety protection threshold is triggered, and store the abnormal state transition data in a preset experience playback pool for offline retraining of the reinforcement learning model of temperature and humidity coupling prediction model and fusion model prediction control.

[0052] In the aforementioned safe grid connection steps, independent of the cloud algorithm network, the programmable logic controller (PLC) within the local control device uses high-priority hardware interrupts to instantly compare the unfiltered current actual temperature and humidity data with the hard safety protection threshold ranges of the physical process. If a hard limit is exceeded, it indicates a sudden extreme condition such as large-scale damage to doors and windows. At this point, the local hardware directly cuts off the cloud command channel, suspending the execution of the control strategy.

[0053] While cutting off the channel, the local device seamlessly activates a conservative control strategy based on the classic Proportional Integral Derivative (PID) algorithm. This algorithm does not rely on future predictions but only applies force according to the current absolute deviation ratio, and combines the integral accumulation of the deviation to forcibly suppress oscillations, using the fastest physical output priority to force environmental parameters back to the safe baseline.

[0054] After an incident, multi-dimensional sensor data before and after the trigger threshold, along with the action sequences issued at that time, are packaged into high-value anomalous state transition data. This data is stored in a pre-defined experience replay pool. At night or during periods of server idle computing power, cloud-based algorithms replay these extreme edge data that led to failures into the temperature and humidity coupled prediction model and reinforcement learning model, using gradient backpropagation to adjust the weights and retrain the neural networks offline. Through this mechanism, the system's prediction fault tolerance and policy robustness will continuously evolve.

[0055] This invention also provides an artificial intelligence-based energy-saving control device for constant temperature and humidity air conditioning, referring to... Figure 3 The diagram shows a functional block diagram of an artificial intelligence-based constant temperature and humidity air conditioning energy-saving control device 300 of the present invention. The device may include the following modules: Data acquisition module 301 is used to collect environmental data of constant temperature and humidity space and operating status data of air conditioning equipment; The dynamic prediction module 302 is used to input environmental data and operating status data into the pre-trained temperature and humidity coupled prediction model and energy consumption prediction model, and output the dynamic prediction trend of temperature, dynamic prediction trend of humidity and energy consumption change trend in the future control cycle, respectively. The boundary constraint module 303 is used to calculate the temperature and humidity prediction uncertainty based on the temperature dynamic prediction trend and the humidity dynamic prediction trend, and tighten the preset control boundary parameters when the temperature and humidity prediction uncertainty is higher than the preset uncertainty threshold. The objective construction module 304 is used to construct a multi-objective function under the constraints of control boundary parameters; The strategy generation module 305 is used to generate regulation strategies based on a reinforcement learning model that employs fusion model predictive control, using a multi-objective function and energy consumption change trend. The hierarchical control module 306 is used to allocate the control strategy to the corresponding actuators for hierarchical control based on the response time of the multiple actuators contained in the air conditioning equipment.

[0056] Based on the same inventive concept, another embodiment of the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus. Memory, used to store computer programs; When the processor executes the program stored in the memory, it implements the artificial intelligence-based energy-saving control method for constant temperature and humidity air conditioning of the present invention.

[0057] The communication bus mentioned in the aforementioned electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the figure, but this does not indicate that there is only one bus or one type of bus. The communication interface is used for communication between the aforementioned electronic device and other devices. The memory can include RAM, or it can include NVM, such as at least one disk storage device. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor.

[0058] The processors mentioned above can be general-purpose processors, including CPUs, network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0059] Furthermore, to achieve the above objectives, embodiments of the present invention also propose a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the AI-based energy-saving control method for constant temperature and humidity air conditioning according to embodiments of the present invention.

[0060] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products embodied on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0061] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, electronic devices, apparatuses, and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0062] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0063] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0064] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. "And / or" indicates that either one or both can be chosen. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes the element.

[0065] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An energy-saving control method for constant temperature and humidity air conditioning based on artificial intelligence, characterized in that, include: Collect environmental data of the constant temperature and humidity space and the operating status data of the air conditioning equipment; The environmental data and the operating status data are input into the pre-trained temperature and humidity coupled prediction model and energy consumption prediction model, respectively, and the dynamic prediction trends of temperature, humidity and energy consumption in the future control cycle are output. Based on the dynamic temperature prediction trend and the dynamic humidity prediction trend, the temperature and humidity prediction uncertainty is calculated, and when the temperature and humidity prediction uncertainty is higher than the preset uncertainty threshold, the preset control boundary parameters are tightened. Under the constraints of the control boundary parameters, a multi-objective function is constructed; Based on the multi-objective function and the energy consumption change trend, a reinforcement learning model with fusion model predictive control is used to generate a regulation strategy. Based on the response time of the multiple actuators included in the air conditioning equipment, the control strategy is allocated to the corresponding actuators for hierarchical control.

2. The energy-saving control method for constant temperature and humidity air conditioning based on artificial intelligence according to claim 1, characterized in that, The collection of environmental data from the constant temperature and humidity space and the operating status data of the air conditioning equipment includes: The constant temperature and humidity space is divided into multiple independent control areas; The indoor temperature, outdoor temperature, relative humidity, dew point temperature, supply and return air status, space load, personnel heat generation, and process equipment heat generation of each control area are collected in real time as environmental data. The operating power, compressor frequency, fan frequency, chilled water valve opening, humidification status, dehumidification status, and reheat status of each of the multiple actuators are collected synchronously as the operating status data. The environmental data and the operational status data are uploaded to a cloud server, so that the regulation strategy can be remotely generated through the reinforcement learning model of the temperature and humidity coupling prediction model, the energy consumption prediction model and the fusion model prediction control within the cloud server.

3. The energy-saving control method for constant temperature and humidity air conditioning based on artificial intelligence according to claim 1, characterized in that, The step of inputting the environmental data and the operating status data into a pre-trained temperature and humidity coupled prediction model and an energy consumption prediction model, and outputting the dynamic prediction trends of temperature, humidity, and energy consumption within the future control cycle, respectively, includes: The environmental data and the operational status data are cleaned and feature standardized to construct a fusion feature vector that includes environmental status features, equipment action features, and historical response features. The fused feature vector is input into the temperature and humidity coupled prediction model. The graph neural network in the temperature and humidity coupled prediction model is used to extract the spatial thermodynamic interaction correlation features in the constant temperature and humidity space. The long short-term memory network is used to extract the dynamic temporal dependence features of the temperature and humidity over time. The spatial thermal interaction correlation features and the dynamic temporal dependence features are weighted and fused through an attention mechanism, and a nonlinear coupling dynamic response relationship between temperature and humidity is established through a latent variable sharing layer. The output includes the temperature dynamic prediction trend containing the temperature mean and temperature variance, and the humidity dynamic prediction trend containing the humidity mean and humidity variance. The fused feature vector is input into the energy consumption prediction model, and the correlation between device state changes and energy consumption is mapped through a multilayer perceptron network to output the energy consumption change trend containing the expected power consumption.

4. The energy-saving control method for constant temperature and humidity air conditioning based on artificial intelligence according to claim 1, characterized in that, The step of calculating the temperature and humidity prediction uncertainty based on the dynamic prediction trends of temperature and humidity, and tightening the preset control boundary parameters when the temperature and humidity prediction uncertainty exceeds a preset uncertainty threshold, includes: From the dynamic prediction trends of temperature and humidity, a prediction variance matrix containing the prediction variance is extracted, and the trace of the prediction variance matrix is ​​used as the uncertainty of the temperature and humidity prediction to characterize the confidence of the model prediction. Determine whether the uncertainty of the temperature and humidity prediction is higher than the uncertainty threshold; When the uncertainty of the temperature and humidity prediction exceeds the uncertainty threshold, the current system state is determined to be in a data-sparse dynamic edge condition, triggering the limit-breaking risk penalty mechanism. The boundary contraction coefficient is calculated based on the ratio of the difference between the temperature and humidity prediction uncertainty and the uncertainty threshold. The upper and lower boundaries of the preset temperature tolerance range and humidity tolerance range in the control boundary parameters are proportionally compressed using the boundary contraction coefficient to obtain the tightened control boundary parameters. When the uncertainty of the temperature and humidity prediction is not higher than the uncertainty threshold, the initial temperature tolerance range and humidity tolerance range are maintained as the control boundary parameters.

5. The energy-saving control method for constant temperature and humidity air conditioning based on artificial intelligence according to claim 1, characterized in that, The construction of the multi-objective function, and the generation of a regulation strategy using a reinforcement learning model based on the multi-objective function and the energy consumption change trend, employing a fusion model predictive control, include: The multi-objective function is constructed with the objectives of minimizing temperature deviation, minimizing humidity deviation, minimizing energy consumption, and minimizing the number of equipment operations. The dynamic temperature prediction trend, the dynamic humidity prediction trend, the energy consumption change trend, and the control boundary parameters are used as the environmental state space and input into the reinforcement learning model of the fusion model predictive control. Starting from the current environmental state space, the model prediction path integral algorithm is used to introduce Gaussian noise for random sampling within a preset continuous action space to generate a preset number of candidate action sequences. The comprehensive cost value of each set of candidate action sequences in the future control cycle is calculated using the multi-objective function. The comprehensive cost value is obtained by weighted summation of temperature deviation cost, humidity deviation cost, energy consumption prediction cost, and equipment operation frequency penalty cost. When the temperature and humidity values ​​predicted by the candidate action sequence exceed the control boundary parameters, the corresponding temperature deviation cost and humidity deviation cost are given a step penalty weight. Based on the comprehensive cost value of all groups of candidate action sequences, calculate the information theory probability weight of each group of candidate action sequences; The candidate action sequences of all groups are weighted and averaged using the information theory probability weights, and then fused to generate the globally optimal action sequence. Extract the set of optimal actions corresponding to the first control cycle in the global optimal action sequence to generate the control strategy containing multiple control actions.

6. The energy-saving control method for constant temperature and humidity air conditioning based on artificial intelligence according to claim 5, characterized in that, The step of allocating the control strategy to the corresponding actuators for hierarchical control based on the response time of the multiple actuators included in the air conditioning equipment includes: Based on the mechanical inertia, thermodynamic response delay time, and control constraint complexity of each of the plurality of actuators, the plurality of actuators are divided into a slow response layer, a medium response layer, and a fast response layer. The slow response layer is configured to include a compressor start / stop control component for providing the basic cold and heat source and a chilled water main valve regulating component, and the control frequency of the slow response layer is set to a first cycle; The medium-speed response layer is configured to include a fan frequency adjustment component and a reheat compensation device for coarse adjustment of sensible heat load and latent heat load, and the control frequency of the medium-speed response layer is set to a second cycle that is less than the first cycle. The fast response layer is configured to include a humidification device and a dehumidification device for precise temperature and humidity fine-tuning, and the control frequency of the fast response layer is set to a third cycle that is less than the second cycle. Using a hierarchical reinforcement learning framework, the control strategy is decomposed into a first control instruction decided by a high-level agent, a second control instruction decided by a mid-level agent, and a third control instruction decided by a low-level agent according to the time scale. The first control command is sent to the slow response layer, the second control command is sent to the medium response layer, and the third control command is sent to the fast response layer, so that the execution components with different response speeds can perform actions asynchronously within their respective cycles to complete the hierarchical control.

7. The energy-saving control method for constant temperature and humidity air conditioning based on artificial intelligence according to claim 6, characterized in that, The method further includes: During the execution of the hierarchical control, the current actual temperature and humidity data of the constant temperature and humidity space are obtained in real time by continuously collecting the environmental data. The current actual temperature data and the current actual humidity data are compared with the hard safety protection threshold range that is pre-set for the constant temperature and humidity space; If the current actual temperature data or the current actual humidity data exceeds the hard safety protection threshold range, it is determined that the current system has encountered a sudden strong disturbance or the control strategy has a risk of exceeding the limit, and the execution of the control strategy generated by the reinforcement learning model of the fusion model predictive control is immediately suspended. The control mode is automatically switched to a conservative control strategy based on the proportional-integral-derivative algorithm to prioritize restoring the temperature and humidity of the constant temperature and humidity space to within the safe baseline. Collect abnormal state transition data at the moment when the hard safety protection threshold range is triggered, and store the abnormal state transition data in a preset experience playback pool for offline retraining of the reinforcement learning model of the temperature and humidity coupled prediction model and the fusion model prediction control.

8. An energy-saving control device for constant temperature and humidity air conditioning based on artificial intelligence, characterized in that, include: The data acquisition module is used to collect environmental data of the constant temperature and humidity space and the operating status data of the air conditioning equipment; The dynamic prediction module is used to input the environmental data and the operating status data into the pre-trained temperature and humidity coupled prediction model and energy consumption prediction model, and output the dynamic prediction trend of temperature, dynamic prediction trend of humidity and energy consumption change trend in the future control cycle, respectively. The boundary constraint module is used to calculate the temperature and humidity prediction uncertainty based on the temperature dynamic prediction trend and the humidity dynamic prediction trend, and tighten the preset control boundary parameters when the temperature and humidity prediction uncertainty is higher than the preset uncertainty threshold. The objective construction module is used to construct a multi-objective function under the constraints of the control boundary parameters; The strategy generation module is used to generate a regulation strategy based on the multi-objective function and the energy consumption change trend, using a reinforcement learning model of fusion model predictive control. The hierarchical control module is used to allocate the control strategy to the corresponding actuators for hierarchical control based on the response time of the multiple actuators included in the air conditioning equipment.

9. An electronic device, characterized in that, include: At least one processor; The system includes a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, and when the computer program is executed by the at least one processor, it implements the artificial intelligence-based energy-saving control method for constant temperature and humidity air conditioning as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the artificial intelligence-based energy-saving control method for constant temperature and humidity air conditioning as described in any one of claims 1 to 7.

Images