Control method of air conditioning system and air conditioning system

Abstract

The present application relates to the technical field of air conditioning equipment, and particularly provides a control method of an air conditioning system and the air conditioning system, aiming to solve the problems of high development cost, long development period, high storage and maintenance cost of the intelligent optimization control technology based on historical data modeling, which needs to train a special control model. To this end, the control method of the air conditioning system comprises the following steps: training a general model according to a sample data set of a sample air conditioning system; deploying the general model to form an edge model by the air conditioning system; and determining an air conditioning operation strategy according to a real-time data set of the air conditioning system, the edge model and a control algorithm. The air conditioning system of the present application forms an edge model by deploying the trained general model, so that the edge model does not need to be modeled from scratch, and the air conditioning system can determine the air conditioning operation strategy according to the edge model, the real-time data set and the control algorithm, thereby reducing the development, storage and maintenance costs and shortening the development period.

Description

Technical Field

[0001] This invention relates to the field of air conditioning equipment technology, specifically providing a control method for an air conditioning system and an air conditioning system. Background Technology

[0002] In recent years, with the increasing requirements for building energy conservation and the rise of artificial intelligence and big data technologies, intelligent control of HVAC systems has become a key research focus in the industry. Its core objective is to maintain indoor thermal comfort while minimizing system energy consumption, and to balance temperature stability control and energy efficiency optimization.

[0003] In existing technologies, intelligent optimization control technology based on historical data modeling is commonly used to control the operation of air conditioning systems. This type of method can build a system model by learning from a large amount of system operating data, achieving multi-objective synergistic optimization of energy efficiency and comfort. However, in practical applications, it is necessary to collect historical operating data under all operating conditions separately for each specific project and different equipment system, and use this data to train a dedicated control model. This results in high overall development costs and a long development cycle. Furthermore, as the number of equipment models and projects increases, storage and maintenance costs also increase.

[0004] Accordingly, there is a need in the field for a new control method and air conditioning system for air conditioning systems to solve the above problems. Summary of the Invention

[0005] The present invention aims to solve the above-mentioned technical problems, namely, to solve the problem that in the practical application of intelligent optimization control technology based on historical data modeling, it is necessary to collect the full-condition historical operating data of each specific project and different equipment system separately, and use it to train a dedicated control model, which leads to high overall development costs, long cycle, and increased storage and maintenance costs as the number of equipment models and projects increases.

[0006] In a first aspect, the present invention provides a control method for an air conditioning system; the control method for the air conditioning system includes the following steps: A general model was trained based on the sample dataset of the sample air conditioning system. The air conditioning system deploys the general model to form an edge model; The air conditioning operation strategy is determined based on the real-time dataset of the air conditioning system, the edge model, and the control algorithm.

[0007] In the preferred embodiment of the above control method, the air conditioning system is equipped with conditions for triggering adaptive updates; When the adaptive update condition is triggered, the edge model is incrementally updated based on the historical dataset; The air conditioning operation strategy is determined based on the real-time dataset, the updated edge model, and the control algorithm.

[0008] In the preferred technical solution of the above control method, "training a general model based on the sample dataset of the sample air conditioning system" specifically includes: The sample dataset includes a model identifier, a first actual state vector, and a first actual action vector; The model identifier, first actual state vector and first actual action vector of each sample air conditioning system are input into the general model, and the general model outputs the first predicted state vector, the first predicted system power and the prediction uncertainty estimate. Calculate the multi-objective loss function value based on the first actual state vector, the first predicted state vector, the first predicted system power, the prediction uncertainty estimate, and the multi-objective loss function set; Backpropagation is performed based on the multi-objective loss function value to update the general model parameters; Based on the multi-objective loss function value and the number of training rounds, determine whether the general model has ended training; Preferably, "determining whether the general model has ended training based on the multi-objective loss function value and the number of training epochs" specifically includes: If the multi-objective loss function value is greater than a preset multi-objective loss function value, and the number of training epochs is less than a preset number of epochs, continue training the general model; and / or If the value of the multi-objective loss function is not greater than the preset multi-objective loss function value, or if the number of training rounds is not less than the preset number of rounds, the training of the general model is complete.

[0009] In the preferred embodiment of the above control method, the multi-objective loss function set includes: in, : Multi-objective loss function value; : Weights of the state prediction loss; State prediction loss; Weights of energy consumption prediction losses; Energy consumption prediction loss; Weights for uncertainty-calibrated prediction losses; Uncertainty calibration loss; : Training sequence length; : First predicted state vector; : First actual state vector; First prediction system power; Actual system power; : Estimation of forecast uncertainty.

[0010] In the preferred technical solution of the above control method, the general model includes a feature extraction layer, a dynamic prediction layer and a multi-target output layer. The feature extraction layer can embed and encode the model identifier, extract features from the first actual state vector and the first actual action vector, and concatenate the extracted features to obtain the concatenated input feature quantity. The dynamic prediction layer can generate features output by spatial attention mechanism after passing the concatenated input features through a multi-head attention mechanism network, and can also generate feature vectors output by temporal gated recurrent units after passing the features output by spatial attention mechanism through a bidirectional gated recurrent network. The multi-objective output layer can output a first predicted state vector, a first predicted system power, and a prediction uncertainty estimate based on the feature vector output by the time-gated loop unit. Preferably, the feature extraction layer includes: in, Model number embedded vector, dimension 64; Embedded layer, containing model identifier It is mapped to a 64-dimensional vector; : Identifier of the i-th type of system equipment; : The feature vector after state encoding, with a dimension of 64; : The weight matrix of the state coding layer; : The bias vector of the state coding layer; : Feature vector after action encoding, dimension 32; : Weight matrix of the action coding layer; : Bias vector of the action coding layer; : The concatenated input features, dimension 160; Preferably, the dynamic prediction layer includes: in, Features of the output of spatial attention mechanisms; Multi-head attention mechanism; : Feature vector output by temporal GRU, dimension 512; Two-way gated loop unit; Preferably, the multi-target output layer includes: in, : First predicted state vector; Weights for state prediction; : Bias in state prediction; First prediction system power; Weighting of energy consumption forecasts; : Bias in energy consumption prediction; : Estimation of uncertainty in forecasting; Weights for uncertainty estimation; Bias in uncertainty estimation.

[0011] In the preferred embodiment of the above control method, the adaptive update conditions include the average room temperature control accuracy being lower than a first preset value, the system energy efficiency being lower than a second preset value, or the compressor frequency control stability being lower than a third preset value; and / or The average temperature control accuracy of all rooms Calculated using the formula: in, Number of rooms; The current indoor temperature of the i-th room; The set temperature of the i-th room; and / or The system energy efficiency Calculated using the formula: in, Total cooling / heating capacity of the system; System power; and / or The compressor frequency control stability S is calculated using the following formula: in, : Change in compressor frequency; Variance calculation; : The maximum allowable frequency variation of the compressor.

[0012] In the preferred embodiment of the above control method, "incrementally updating the edge model based on the historical dataset" specifically includes: The historical dataset includes the pre-trigger window dataset and the experience replay buffer dataset; Create a copy of the edge model; The operation of the air conditioning system is controlled according to the edge model. At the same time, the pre-trigger window dataset and the experience replay buffer dataset are input into the edge model replica for forward propagation. The total adaptive learning loss is calculated based on the current edge model replica parameters, edge model parameters and protective loss function set. Backpropagation is performed based on the total adaptive learning loss to update the parameters of the edge model replicas; Based on the total adaptive learning loss and the number of gradient update iterations, determine whether to update the edge model; Preferably, "determining whether to update the edge model based on the total adaptive learning loss and the number of gradient update iterations" specifically includes: If the total adaptive learning loss is not greater than a preset total adaptive learning loss, or the number of gradient update iterations is not less than a preset number of iterations, update the edge model; and / or If the total adaptive learning loss is greater than the preset total adaptive learning loss, and the number of gradient update iterations is less than the preset number of iterations, the pre-trigger window dataset and the experience replay buffer dataset are input into the edge model replica for forward propagation, and the total adaptive learning loss is calculated based on the current edge model replica parameters, edge model parameters, and protective loss function set.

[0013] In the preferred embodiment of the above control method, the pre-trigger window dataset includes a second actual state vector and a second actual action vector. The second actual state vector includes a second current state vector and a second next state vector. The second actual action vector includes a second current action vector. The experience replay buffer dataset includes a third actual state vector and a third actual action vector. The third actual state vector includes a third current time state vector and a third next time state vector. The third actual state vector includes a third current time action vector. The protective loss function set includes: in, Total loss in adaptive learning; Local data prediction loss; Elastic weights consolidate loss weights; Elastic weight consolidation loss; : Experience replay loss weight; Loss of experience replay; : Number of local data samples; : The current copy of the edge model; : The second current state vector; : The action vector at the second current moment; : The state vector of the second next time step; : Importance weights; : Current edge model replica parameters; : Current edge model parameters; Number of samples in the experience replay buffer; : The current state vector at the third moment; : The action vector at the current moment; : The state vector at the third next time step.

[0014] In the preferred embodiment of the above control method, "determining the air conditioning operation strategy based on the real-time dataset, the updated edge model, and the control algorithm" specifically includes: The real-time dataset includes a real-time current state vector; The real-time current state vector, the current predictive control action vector sequence, and the model identifier are input into the updated edge model. The updated edge model is called multiple times through recursive calls. The edge model outputs multiple second predictive state vectors and second predictive system power in sequence, forming a second predictive state vector sequence and a second predictive system power sequence. The total objective function value is calculated based on the predictive control action vector sequence, the second predictive state vector sequence, the second predictive system power sequence, and the objective function. Based on the total objective function value and the preset total objective function value, determine whether the control algorithm has converged; If the control algorithm fails to converge, the gradient is calculated by backpropagation based on the total objective function value, the predicted control action vector sequence, the second predicted state vector sequence, and the solution function, to obtain the predicted control action vector sequence. Based on the control action amplitude constraint, the predicted control action vector sequence is processed to obtain the updated predicted control action vector sequence; The real-time current state vector, the updated predictive control action vector sequence, and the model identifier are input into the updated edge model. The updated edge model is then called multiple times via recursive calls. The edge model sequentially outputs multiple second predictive state vectors and second predictive system power, forming a second predictive state vector sequence and a second predictive system power sequence; and / or If the control algorithm converges, the predicted control action vector sequence corresponding to the current total objective function value is subjected to safety constraint processing according to the safety constraint function set to obtain the optimal predicted control action vector sequence. The air conditioning operation strategy is executed, which is the first step in executing the optimal predictive control action vector sequence; and / or The objective function is: in, : Overall objective function value; Predict the length of the time series; : The sequence of control actions from time t to t+H-1; Temperature tracking weights; The predicted temperature at time t+k+1; Set the temperature; Energy consumption optimization weights; Predicted system power at time t+k+1; : Control the smoothness weight; : The control action vector at time t+k; and / or The solution function: in, The overall objective function value; The gradient of the objective function with respect to the control action; The gradient of the objective function with respect to the state; : The Jacobian matrix of state to control action.

[0015] In a second aspect, the present invention provides an air conditioning system; the air conditioning system includes a control module configured to execute the control method described above.

[0016] By employing the aforementioned technical solution, a general model is trained using sample datasets from multiple models and projects, enabling the general model to adapt to applications of different air conditioning systems. The trained general model is deployed on the air conditioning system, forming an edge model. Deploying the general model achieves cross-device and cross-project knowledge reuse, eliminating the need for the edge model to be built from scratch. The air conditioning system can determine its operating strategy based on the edge model, real-time dataset, and control algorithm, thereby reducing the development cost and time required for building the edge model from scratch. Furthermore, in existing technologies, as the number of air conditioning systems increases, the number of edge models also increases, and the amount of data required for building from scratch also increases. Compared to existing technologies, in this invention, when there are multiple air conditioning systems, each air conditioning system deploys a general model instead of training a dedicated edge model from scratch, reducing data storage costs.

[0017] Furthermore, when the air conditioning system triggers the adaptive update condition, the edge model is incrementally updated based on the historical dataset. The air conditioning operation strategy is determined based on the historical dataset, the updated edge model, and the control algorithm. Updating the edge model only when the adaptive update condition is triggered allows for timely updates, making it more suitable for the actual operating conditions of the air conditioning system compared to periodic updates. This also reduces unnecessary computation during frequent updates and lowers the computational cost of the edge model. Moreover, the incremental updates of the edge model, with protective loss functions including local data prediction loss, elastic weight consolidation loss, and experience replay loss, enable the edge model to learn the actual operating conditions of the air conditioning system while effectively preventing catastrophic forgetting of common mechanisms learned from the general model. This makes the edge model more suitable for the operating conditions of the air conditioning system, and its suitability increases with operating time. Attached Figure Description

[0018] The preferred embodiments of the present invention are described below with reference to the accompanying drawings, in which: Figure 1 This is a flowchart of the main steps of the control method for the air conditioning system of the present invention; Figure 2 This is a flowchart illustrating the specific steps of the control method for the air conditioning system of the present invention; Figure 3 yes Figure 2 The detailed flowchart of step S1; Figure 4 yes Figure 3 Flowchart of the specific steps in step S15; Figure 5 yes Figure 2 Flowchart of the specific steps in step S6; Figure 6 yes Figure 2 The flowchart of the specific steps in step S7. Detailed Implementation

[0019] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the invention and are not intended to limit the scope of protection of the invention. Those skilled in the art can make adjustments as needed to adapt to specific applications.

[0020] It should be noted that in the description of this invention, terms such as "upper," "lower," "left," and "right," indicating directional or positional relationships, are based on the directional or positional relationships shown in the accompanying drawings. These are merely for ease of description and do not indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation; therefore, they should not be construed as limitations on this invention. Furthermore, the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0021] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "connected," "linked," and "connected" should be interpreted broadly. For example, they can refer to a fixed connection or a detachable connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0022] To address the issue that intelligent optimization control technology based on historical data modeling requires the separate collection of full-condition historical operating data for each specific project and different equipment system model in practical applications, and the use of this data to train a dedicated control model, the overall development costs are high and the cycle is long. Furthermore, as the number of equipment models and projects increases, the storage and maintenance costs also increase.

[0023] like Figures 1 to 6 As shown, this embodiment discloses a control method for an air conditioning system. The air conditioning system is a heating, ventilation, and air conditioning system, specifically including a multi-split air conditioning system, a heat pump, a chiller unit, or a combined air conditioning system. In this embodiment, a multi-split air conditioning system is used as an example for explanation. A multi-split air conditioning system includes an outdoor unit connected via refrigerant circulation piping and multiple indoor units connected in parallel. The outdoor unit includes a compressor, and adjusting the compressor frequency can change the cooling or heating capacity output of the air conditioning system. The indoor units include electronic expansion valves, and adjusting the opening degree of the electronic expansion valves can control the refrigerant flow into the corresponding indoor unit. By adjusting the opening degree of each electronic expansion valve, refrigerant can be distributed on demand, allowing each room to reach its set temperature.

[0024] The multi-split air conditioning system also includes a control module, which is configured to perform the control methods described below.

[0025] like Figure 1 As shown, the control method of the air conditioning system mainly includes the following steps: S1. Based on the sample dataset of the sample air conditioning system, a general model is trained.

[0026] S2, Air Conditioning System Deployment General Model to Form Edge Model.

[0027] S3. Determine the air conditioning operation strategy based on the real-time dataset, edge model, and control algorithm of the air conditioning system.

[0028] By training a general model using sample datasets from multiple models and projects, this general model can be adapted to applications of different air conditioning systems. The trained general model is deployed on the air conditioning system, forming an edge model. Deploying the general model enables knowledge reuse across devices and projects, eliminating the need for the edge model to be built from scratch. The air conditioning system can determine its operating strategy based on the edge model, real-time dataset, and control algorithm, thereby reducing the development cost and time required for building edge models from scratch. Furthermore, in existing technologies, as the number of air conditioning systems increases, the number of edge models also increases, as does the amount of data required for building models from scratch. Compared to existing technologies, this invention deploys a general model on each of the multiple air conditioning systems, replacing the need to train a dedicated edge model from scratch, thus reducing data storage costs.

[0029] like Figure 2 As shown, the control method of the air conditioning system mainly includes the following steps: S1. A general model is trained based on the sample dataset of the sample air conditioning systems. The sample air conditioning systems include multiple air conditioning systems of different models and projects. The sample dataset comes from laboratory simulation data and actual operation data of multiple air conditioning systems, thus providing sufficient sample data of various models and quantities for training the general model and improving the applicability of the general model.

[0030] S2, Air Conditioning System Deployment General Model to Form Edge Model.

[0031] S3. Obtain the real-time dataset of the air conditioning system. The real-time dataset includes the real-time current state vector, the real-time current action vector, and the model identifier. The real-time current state vector includes the current indoor temperature of each room, the set temperature of each room, the current opening degree of each electronic expansion valve, the outdoor ambient temperature, the compressor frequency, and the system power. The real-time current action vector includes the change in compressor frequency and the change in the opening degree of each electronic expansion valve.

[0032] S4. Determine the air conditioning operation strategy based on the real-time dataset, edge model, and control algorithm of the air conditioning system. A general model is trained using sample datasets from multiple models and projects, enabling it to adapt to different air conditioning system applications. The trained general model is deployed on the local air conditioning system, forming the edge model. Deploying the general model achieves cross-device and cross-project knowledge reuse, eliminating the need for the edge model to be built from scratch. The air conditioning system can determine its operation strategy based on the edge model, real-time dataset, and control algorithm, thus reducing the development cost and cycle of building the edge model from scratch. Furthermore, it allows the air conditioning system to achieve good control performance from the initial operation stage, realizing multi-objective collaborative optimization of temperature stability control and energy-saving control.

[0033] Furthermore, in existing technologies, as the number of air conditioning systems increases, the number of edge models also increases, and the amount of data to be modeled from scratch also increases. Compared to existing technologies, in this invention, when there are multiple air conditioning systems, each air conditioning system deploys a common model instead of training a proprietary edge model from scratch, which can reduce data storage costs. The more air conditioning systems that share a common model, the greater the total savings in development, storage, and maintenance costs, and the greater the total savings in development time.

[0034] S5. Determine whether the air conditioning system has triggered the adaptive update condition. The adaptive update condition includes the average temperature control accuracy of all rooms being lower than the first preset value, the system energy efficiency being lower than the second preset value, or the compressor frequency control stability being lower than the third preset value. If the adaptive update condition is triggered, execute S6. If the adaptive update condition is not triggered, return to execute S3 after the first control cycle is reached.

[0035] Specifically, based on the real-time current state vector, the real-time current action vector, and formulas (17) to (19), the average room temperature control accuracy, system energy efficiency, and compressor frequency control stability are calculated. If the average room temperature control accuracy is lower than the first preset value, the system energy efficiency is lower than the second preset value, or the compressor frequency control stability is lower than the third preset value, the air conditioning system triggers the adaptive update condition and executes S5; otherwise, after reaching the first control cycle, it returns to execute S3.

[0036] Average temperature control accuracy in all rooms Calculated using the formula: in, Number of rooms; The current indoor temperature of the i-th room; : The set temperature of the i-th room.

[0037] System energy efficiency Calculated using the formula: in, Total cooling / heating capacity of the system; System power.

[0038] The compressor frequency control stability S is calculated using the formula: in, : Change in compressor frequency; Variance calculation; : The maximum allowable frequency variation of the compressor.

[0039] The air conditioning system updates the edge model only when the adaptive update condition is triggered. Compared with periodic updates, the edge model can be updated in a timely manner, making it more suitable for the actual operating conditions of the air conditioning system. At the same time, it reduces invalid calculations during frequent updates and lowers the computing power consumption of the edge model.

[0040] S6. Based on the historical dataset, incrementally update the edge model. The historical dataset includes the model identifier, the pre-trigger window dataset, and the experience replay buffer dataset. The pre-trigger window dataset contains all running data within a preset time period before triggering the adaptive update. The experience replay buffer dataset contains randomly sampled data from multiple different time periods and operating conditions in all historical running data.

[0041] The edge model is incrementally updated based on the air conditioning system's model identifier, the pre-trigger window dataset, and the experience replay buffer dataset. This allows the edge model to learn the actual operating conditions of the air conditioning system, improving its adaptability to the application scenario. It also effectively prevents the catastrophic forgetting of common mechanism knowledge learned from the general model. As a result, the edge model becomes more suitable for the operating conditions of the air conditioning system, and it becomes increasingly suitable for the operating conditions of the air conditioning system as the operating time increases.

[0042] S7. Based on the real-time dataset and the updated edge model, the control algorithm determines the air conditioning operation strategy. S4 and S7 have the same specific steps, the only difference being that S4 calls the unupdated edge model (which is also a general model), while S6 calls the updated edge model. In this embodiment, step S6 is used as an example for specific explanation. During the operation of the air conditioning system, the control algorithm can comprehensively consider multiple objectives such as temperature comfort, system power minimization, and operational stability. Under the premise of physical safety boundaries such as compressor frequency and electronic expansion valve opening, it predicts future states through the edge model and dynamically generates control strategies using a rolling optimization method. This makes the control algorithm not only predictive, able to respond to load changes in advance, but also adaptable to different equipment models, achieving the goals of temperature comfort, low system power, and stable operation while ensuring the safety and reliability of the air conditioning system.

[0043] S8. Determine again whether the air conditioning system triggers the adaptive update condition. If the adaptive update condition is triggered, return to step S6 and update the current edge model. If the adaptive update condition is not triggered, after the second control cycle is reached, return to step S7 and continue using the current edge model.

[0044] like Figure 3 As shown, Figure 2 Step S1 specifically includes: S11. Obtain the sample dataset, which includes the model identifier, the first actual state vector, and the first actual action vector.

[0045] The first actual state vector includes the current indoor temperature of each room at each time point, the set temperature of each room, the current opening degree of each electronic expansion valve, the outdoor ambient temperature, the compressor frequency, and the system power, as shown in formula (1): in, Current indoor temperature in each room; : The set temperature of each room; : The current opening degree of each electronic expansion valve; Outdoor ambient temperature; Compressor frequency; System power.

[0046] The first actual action vector includes the compressor frequency change and the opening degree change of each electronic expansion valve, specifically as shown in formula (2): in, : Change in compressor frequency; : The change in opening degree of each electronic expansion valve.

[0047] Model designation is shown in formula (3): in, : Identifier of the i-th type of system equipment.

[0048] S12. Generate state and action training samples from the sample dataset using the time series.

[0049] S13. Preprocess the state and action training samples, including anomaly detection, anomaly handling, normalization, and sequence segmentation, to ensure data quality and the stability of general model training.

[0050] S14. Generate batch training data. Due to the large amount of data in the sample dataset, batch training can reduce the computational requirements of the general-purpose model. If the amount of data is small or the computational power of the general-purpose model is sufficient, batch training may not be necessary.

[0051] S15. General model training: During training, a general model is trained on each batch of training data, and a general model is obtained after training.

[0052] During general model training, the model identifier, first actual state vector, and first actual action vector of each sample air conditioning system are input into the general model. The general model outputs a first predicted state vector, a first predicted system power, and a prediction uncertainty estimate. Based on the first actual state vector, the first predicted state vector, the first predicted system power, the prediction uncertainty estimate, and the multi-objective loss function set, the multi-objective loss function value is calculated. Backpropagation is performed based on the multi-objective loss function value to update the general model parameters. The training cycle and the multi-objective loss function value determine whether the general model training has ended.

[0053] The general model includes a feature extraction layer, a dynamic prediction layer, and a multi-objective output layer, such as... Figure 4 As shown, Figure 2 Step S15, the general model training, specifically includes: S151. The feature extraction layer embeds and encodes the model identifier, extracts features from the first actual state vector and the first actual action vector, and concatenates the extracted features to obtain the concatenated input feature quantity.

[0054] Specifically, the first actual state vector, the first actual action vector, and the model identifier are input for each time step. The concatenated input feature quantities are output through formulas (4) to (7): in, Model number embedded vector, dimension 64; Embedded layer, containing model identifier It is mapped to a 64-dimensional vector; : Identifier of the i-th type of system equipment; : The feature vector after state encoding, with a dimension of 64; : The weight matrix of the state coding layer; : The bias vector of the state coding layer; : Feature vector after action encoding, dimension 32; : Weight matrix of the action coding layer; : Bias vector of the action coding layer; : The concatenated input features, dimension 160.

[0055] S152. The dynamic prediction layer passes the concatenated input features through a multi-head attention network to generate features output by a spatial attention mechanism. The features output by the spatial attention mechanism are then passed through a bidirectional gated recurrent network to generate feature vectors output by a temporally gated recurrent unit.

[0056] Specifically, the input features after concatenation are used to output the feature vectors of the time-gated cyclic unit through formulas (8) and (9).

[0057] in, Features of the output of spatial attention mechanisms; Multi-head attention mechanism; : Feature vector output by temporal GRU, dimension 512; : Two-way gated loop unit.

[0058] S153. The multi-objective output layer outputs a first predicted state vector, a first predicted system power, and a prediction uncertainty estimate based on the feature vector output by the time-gated recurrent unit. The first actual state vector includes system power, and the first predicted state vector also includes system power. However, for the sake of convenience in subsequent calculations, system power is output as a separate term: the first predicted system power.

[0059] Specifically, the feature vector output by the input timing-gated recurrent unit is used to output the first predicted state vector, the first predicted system power, and the prediction uncertainty estimate through formulas (10) to (12): in, : First predicted state vector; Weights for state prediction; : Bias in state prediction; First prediction system power; Weighting of energy consumption forecasts; : Bias in energy consumption prediction; : Estimation of uncertainty in forecasting; Weights for uncertainty estimation; Bias in uncertainty estimation.

[0060] S154. Calculate the multi-objective loss function value based on the first actual state vector, the first predicted state vector, the first predicted system power, the prediction uncertainty estimate, and the multi-objective loss function set.

[0061] Specifically, the multi-objective loss function set includes formulas (13)-(16): in, : Multi-objective loss function value; : Weights of the state prediction loss; State prediction loss; Weights of energy consumption prediction losses; Energy consumption prediction loss; Weights for uncertainty-calibrated prediction losses; Uncertainty calibration loss; : Training sequence length; : First predicted state vector; : First actual state vector; First prediction system power; Actual system power; : Estimation of forecast uncertainty.

[0062] The calculation of the multi-objective loss function includes state prediction loss, energy consumption prediction loss, and uncertainty calibration loss. State prediction loss is related to the actual state vector and the predicted state vector. The state vector, in turn, is related to the current indoor temperature of each room, the set temperature of each room, the outdoor ambient temperature, the current opening degree of each electronic expansion valve, and the compressor frequency. Considering the state vector allows the air conditioning system to take temperature stability control into account. Energy consumption loss takes into account system power, and uncertainty loss takes into account the model and operating conditions of the air conditioning system. By designing the weights for state prediction loss, energy consumption prediction loss, and deterministic calibration prediction loss, a balance can be struck between temperature stability control and energy consumption reduction in the air conditioning system. Furthermore, the general model can be applied to a wide range of models and operating conditions with high accuracy.

[0063] S155. Perform backpropagation based on the multi-objective loss function value to update the general model parameters, which include the weights of state prediction, the biases of state prediction, the weights of energy consumption prediction, the biases of energy consumption prediction, the weights of uncertainty estimation, and the biases of uncertainty estimation.

[0064] S156. Determine whether the multi-objective loss function value is greater than the preset multi-objective loss function value. If it is greater than the preset multi-objective loss function value, execute S157; if it is not greater than the preset multi-objective loss function value, execute S158. The training of the general model is complete.

[0065] S157. Determine if the number of training epochs is less than the preset number of epochs. If it is less than the preset number of epochs, return to S151, at which point the general model parameters have been updated. If it is not less than the preset number of epochs, execute S158, and the general model training is complete.

[0066] like Figure 5 As shown, Figure 2 Step S6 specifically includes: S61. Obtain the historical dataset. The pre-trigger window dataset of the historical dataset includes a second actual state vector and a second actual action vector. The second actual state vector includes a second current state vector and a second next state vector. The second actual action vector includes a second current action vector. The experience replay buffer dataset of the historical dataset includes a third actual state vector and a third actual action vector. The third actual state vector includes a third current state vector and a third next state vector. The third actual state vector includes a third current action vector.

[0067] S62. Create an edge model copy. While updating the edge model copy based on the historical dataset, acquire the real-time dataset. Determine the air conditioning operation strategy based on the real-time dataset, the edge model, and the control algorithm, so that the air conditioning system can still operate stably while the edge model is being updated.

[0068] S63. Input the pre-trigger window dataset and the experience replay buffer dataset into the edge model replica for forward propagation, and calculate the total adaptive learning loss based on the current edge model replica parameters, edge model parameters, and protective loss function set.

[0069] Specifically, the protective loss function set includes: The protective loss function set includes: in, Total loss in adaptive learning; Local data prediction loss; Elastic weights consolidate loss weights; Elastic weight consolidation loss; : Experience replay loss weight; Loss of experience replay; : Number of local data samples; : The current copy of the edge model; : The second current state vector; : The action vector at the second current moment; : The state vector of the second next time step; : Importance weights; : Current edge model replica parameters; : Current edge model parameters; Number of samples in the experience replay buffer; : The current state vector at the third moment; : The action vector at the current moment; : The state vector at the third next time step.

[0070] The edge model is incrementally updated, and the protective loss function during the update includes local data prediction loss, elastic weight consolidation loss and experience replay loss. This allows the edge model to learn the actual operating conditions of the air conditioning system and effectively prevent the catastrophic forgetting of common mechanism knowledge learned from the general model. As a result, the edge model is more suitable for the operating conditions of the air conditioning system, and becomes more and more suitable for the operating conditions of the air conditioning system as the operating time increases.

[0071] S64. Perform backpropagation based on the total adaptive learning loss to update the parameters of the edge model replicas. The edge model replica parameters include the weights of state prediction, the biases of state prediction, the weights of energy consumption prediction, the biases of energy consumption prediction, the weights of uncertainty estimation, and the biases of uncertainty estimation.

[0072] S65. Determine if the total adaptive learning loss is not greater than the preset total adaptive learning loss. If it is not greater than the preset total adaptive learning loss, execute S67 to update the edge model, that is, replace the edge model with an edge model copy. If it is greater than the preset total adaptive learning loss, execute S66.

[0073] S66. Determine if the number of gradient update iterations is not less than the preset number of iterations. If it is not less than the preset number of iterations, proceed to S67 to update the edge model. If it is less than the preset number of iterations, return to step S61. At this point, the edge model replica parameters have been updated, and the edge model continues to be updated.

[0074] like Figure 6 As shown, Figure 2 Step S7 specifically includes: S71. Obtain the real-time dataset, which includes the real-time current state vector and the real-time current action vector.

[0075] S72. Input the real-time current state vector, the current predictive control action vector sequence, and the model identifier into the updated edge model. Call the updated edge model multiple times through recursive calls. The edge model outputs multiple second predictive state vectors and second predictive system power in sequence, forming a second predictive state vector sequence and a second predictive system power sequence.

[0076] Specifically, the current predictive control action vector sequence includes an initial predictive control action vector sequence and an updated predictive control action vector sequence. Within a control cycle, the initial predictive control action vector sequence is used for the first calculation of the total objective function value. This initial sequence is either randomly generated or the optimal predictive control action vector sequence from the previous control cycle. From the second calculation onwards, the updated predictive control action vector sequence is used for subsequent calculations of the total objective function value. This updated sequence can be obtained by solving the function and constraining the control action amplitude.

[0077] The predictive control action vector sequence includes multiple predictive control action vectors arranged in time sequence. The real-time current vector is used as the first state vector, and the first step in the predictive control action vector sequence is used as the first control action vector. According to the system dynamics constraints, formula (25) is used to calculate the second predicted state vector and the second predicted system power. The second predicted state vector includes the second predicted system power, but for ease of subsequent calculation, the second predicted system power is output as a separate item. The updated edge model is called multiple times through recursive calls, and the edge model outputs multiple second predicted state vectors and the second predicted system power in sequence, forming the second predicted state vector sequence and the second predicted system power sequence.

[0078] in, : The state vector at time t+k+1; The updated edge model, in step S4, It is the unupdated edge model, i.e., the general model; : The state vector at time t+k; : Control action vector at time t+k; Model identifier.

[0079] S73. Calculate the total objective function value based on the predictive control action vector sequence, the second predictive state vector sequence, the second predictive system power sequence, and the objective function.

[0080] Specifically, the objective function is: in, : Overall objective function value; Predict the length of the time series; The sequence of control actions from time t to t+H-1 is the predicted control action vector sequence. The initial predicted control action vector sequence is used when initially calculating the total objective function value; subsequent calculations of the total objective function value use the updated initial predicted control action vector sequence. Temperature tracking weights; The predicted temperature at time t+k+1 can be obtained from the second predicted state vector; Set the temperature; Energy consumption optimization weights; The predicted system power at time t+k+1 can be obtained from the second predicted system power. : Control the smoothness weight; The control action vector at time t+k can be obtained from the current predictive control action vector sequence.

[0081] The calculation of the overall objective function value includes temperature, system power, and control actions. By adjusting the predictive control action vector sequence, i.e. selecting the control action vector, the optimal balance between temperature, energy consumption, and control smoothness is found to meet the user's comprehensive needs for comfort, energy saving, and equipment durability.

[0082] S74. Determine whether the total objective function value is less than the preset total objective function value. If it is less than the preset total objective function value, determine that the control algorithm has converged and execute S77; if it is not less than the preset total objective function value, determine that the control algorithm has not converged and execute S75.

[0083] S75. Based on the overall objective function value, the predicted control action vector sequence, the second predicted state vector sequence, and the solution function, the gradient is calculated through backpropagation to obtain the predicted control action vector sequence. Calculating the gradient through backpropagation to obtain the predicted control action vector sequence allows the objective function value to reach the convergence condition (i.e., the overall objective function value is less than the preset overall objective function value) more quickly, enabling the air conditioning system to determine its operating strategy promptly based on the real-time dataset.

[0084] Specifically, the solution function is shown in formula (28).

[0085] in, The overall objective function value; The gradient of the objective function with respect to the control action; The gradient of the objective function with respect to the state; : The Jacobian matrix of state to control action.

[0086] S76. Based on the control action amplitude constraint, process the predicted control action vector sequence to obtain an updated predicted control action vector sequence, and update the predicted control action vector sequence. Then return to step S72, and based on the updated predicted control action vector sequence, continue to form the second predicted state vector sequence and the second predicted system power sequence. By constraining the control action amplitude, the control action can be stably adjusted, reducing the possibility of sudden temperature changes and large changes in compressor frequency caused by over-adjustment.

[0087] Specifically, the control action amplitude constraints are shown in formulas (26) and (27): in, : The state vector at time t+k; : Control action vector at time t+k.

[0088] S77. Based on the safety constraint function set, apply safety constraint processing to the predictive control action vector sequence corresponding to the current total objective function value to obtain the optimal predictive control action vector sequence. Limit the compressor frequency, electronic expansion valve opening, and rate of change using the safety constraint function set to ensure that operation does not exceed hardware limits.

[0089] Specifically, the set of safety constraint functions is shown in formulas (29) to (31): in, Compressor frequency; Compressor frequency lower limit; : Upper limit of compressor frequency; : The opening degree of the i-th electronic expansion valve; Lower limit of valve opening; Valve opening limit: : Change in compressor frequency; : Maximum change in compressor frequency.

[0090] S78. Execute the air conditioning operation strategy. The air conditioning operation strategy is the first step in executing the optimal predictive control action vector sequence.

[0091] During the operation of an air conditioning system, the control algorithm can comprehensively consider multiple objectives such as temperature comfort, system power minimization, and operational stability. Under the premise of physical safety boundaries such as compressor frequency and electronic expansion valve opening, it predicts future states through edge-end models and dynamically generates control strategies using rolling optimization. This enables the control algorithm to not only be predictive, responding in advance to load changes, but also adapt to different equipment models, ensuring the safety and reliability of the air conditioning system while achieving the goals of temperature comfort, low system power consumption, and stable operation.

[0092] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A control method for an air conditioning system, characterized in that, The control method for the air conditioning system includes the following steps: A general model was trained based on the sample dataset of the sample air conditioning system. The air conditioning system deploys the general model to form an edge model; The air conditioning operation strategy is determined based on the real-time dataset of the air conditioning system, the edge model, and the control algorithm.

2. The control method according to claim 1, characterized in that, The air conditioning system is equipped with conditions for triggering adaptive updates. When the adaptive update condition is triggered, the edge model is incrementally updated based on the historical dataset; The air conditioning operation strategy is determined based on the real-time dataset, the updated edge model, and the control algorithm.

3. The control method according to claim 1, characterized in that, "Based on the sample dataset of the sample air conditioning systems, a general model is trained" specifically includes: The sample dataset includes a model identifier, a first actual state vector, and a first actual action vector; The model identifier, first actual state vector and first actual action vector of each sample air conditioning system are input into the general model, and the general model outputs the first predicted state vector, the first predicted system power and the prediction uncertainty estimate. Calculate the multi-objective loss function value based on the first actual state vector, the first predicted state vector, the first predicted system power, the prediction uncertainty estimate, and the multi-objective loss function set; Backpropagation is performed based on the multi-objective loss function value to update the general model parameters; Based on the multi-objective loss function value and the number of training rounds, determine whether the general model has ended training; Preferably, "determining whether the general model has ended training based on the multi-objective loss function value and the training epochs" specifically includes: If the multi-objective loss function value is greater than a preset multi-objective loss function value, and the number of training epochs is less than a preset number of epochs, continue training the general model; and / or If the value of the multi-objective loss function is not greater than the preset multi-objective loss function value, or if the number of training rounds is not less than the preset number of rounds, the training of the general model is complete.

4. The control method according to claim 3, characterized in that, The multi-objective loss function set includes: in, : Multi-objective loss function value; : Weights of the state prediction loss; State prediction loss; Weights of energy consumption prediction losses; Energy consumption prediction loss; Weights for uncertainty-calibrated prediction losses; Uncertainty calibration loss; : Training sequence length; : First predicted state vector; : First actual state vector; First prediction system power; Actual system power; : Estimation of forecast uncertainty.

5. The control method according to claim 3, characterized in that, The general model includes a feature extraction layer, a dynamic prediction layer, and a multi-target output layer. The feature extraction layer can embed and encode the model identifier, extract features from the first actual state vector and the first actual action vector, and concatenate the extracted features to obtain the concatenated input feature quantity. The dynamic prediction layer can generate features output by spatial attention mechanism after passing the concatenated input features through a multi-head attention mechanism network, and can also generate feature vectors output by temporal gated recurrent units after passing the features output by spatial attention mechanism through a bidirectional gated recurrent network. The multi-objective output layer can output a first predicted state vector, a first predicted system power, and a prediction uncertainty estimate based on the feature vector output by the time-gated loop unit. Preferably, the feature extraction layer includes: in, Model number embedded vector, dimension 64; Embedded layer, containing model identifier It is mapped to a 64-dimensional vector; : Identifier of the i-th type of system equipment; : The feature vector after state encoding, with a dimension of 64; : The weight matrix of the state coding layer; : The bias vector of the state coding layer; : Feature vector after action encoding, dimension 32; : Weight matrix of the action coding layer; : Bias vector of the action coding layer; : The concatenated input features, dimension 160; Preferably, the dynamic prediction layer includes: in, Features of the output of spatial attention mechanisms; Multi-head attention mechanism; : Feature vector output by temporal GRU, dimension 512; Two-way gated loop unit; Preferably, the multi-target output layer includes: in, : First predicted state vector; Weights for state prediction; : Bias in state prediction; First prediction system power; Weighting of energy consumption forecasts; : Bias in energy consumption prediction; : Estimation of forecast uncertainty; Weights for uncertainty estimation; Bias in uncertainty estimation.

6. The control method according to claim 2, characterized in that, The adaptive update conditions include the average room temperature control accuracy being lower than the first preset value, the system energy efficiency being lower than the second preset value, or the compressor frequency control stability being lower than the third preset value; and / or The average temperature control accuracy of all rooms Calculated using the formula: in, Number of rooms; The current indoor temperature of the i-th room; The set temperature of the i-th room; and / or The system energy efficiency Calculated using the formula: in, Total cooling / heating capacity of the system; System power; and / or The compressor frequency control stability S is calculated using the following formula: in, : Change in compressor frequency; Variance calculation; : The maximum allowable frequency variation of the compressor.

7. The control method according to claim 2, characterized in that, "Incrementally updating the edge model based on the historical dataset" specifically includes: The historical dataset includes the pre-trigger window dataset and the experience replay buffer dataset; Create a copy of the edge model; The operation of the air conditioning system is controlled according to the edge model. At the same time, the pre-trigger window dataset and the experience replay buffer dataset are input into the edge model replica for forward propagation. The total adaptive learning loss is calculated based on the current edge model replica parameters, edge model parameters and protective loss function set. Backpropagation is performed based on the total adaptive learning loss to update the parameters of the edge model replicas; Based on the total adaptive learning loss and the number of gradient update iterations, determine whether to update the edge model; Preferably, "determining whether to update the edge model based on the total adaptive learning loss and the number of gradient update iterations" specifically includes: If the total adaptive learning loss is not greater than a preset total adaptive learning loss, or the number of gradient update iterations is not less than a preset number of iterations, update the edge model; and / or If the total adaptive learning loss is greater than the preset total adaptive learning loss, and the number of gradient update iterations is less than the preset number of iterations, the pre-trigger window dataset and the experience replay buffer dataset are input into the edge model replica for forward propagation, and the total adaptive learning loss is calculated based on the current edge model replica parameters, edge model parameters, and protective loss function set.

8. The control method according to claim 7, characterized in that, The pre-trigger window dataset includes a second actual state vector and a second actual action vector. The second actual state vector includes a second current state vector and a second next state vector. The second actual action vector includes a second current action vector. The experience replay buffer dataset includes a third actual state vector and a third actual action vector. The third actual state vector includes a third current time state vector and a third next time state vector. The third actual state vector includes a third current time action vector. The protective loss function set includes: in, Total loss in adaptive learning; Local data prediction loss; Elastic weights consolidate loss weights; Elastic weight consolidation loss; : Experience replay loss weight; Loss of experience replay; : Number of local data samples; : The current copy of the edge model; : The second current state vector; : The action vector at the second current moment; : The state vector of the second next time step; : Importance weights; : Current edge model replica parameters; : Current edge model parameters; Number of samples in the experience replay buffer; : The current state vector at the third moment; : The action vector at the third current moment; : The state vector at the third next time step.

9. The control method according to claim 2, characterized in that, "Determining the air conditioning operation strategy based on the real-time dataset, the updated edge model, and the control algorithm" specifically includes: The real-time dataset includes a real-time current state vector; The real-time current state vector, the current predictive control action vector sequence, and the model identifier are input into the updated edge model. The updated edge model is called multiple times through recursive calls. The edge model outputs multiple second predictive state vectors and second predictive system power in sequence, forming a second predictive state vector sequence and a second predictive system power sequence. The total objective function value is calculated based on the predictive control action vector sequence, the second predictive state vector sequence, the second predictive system power sequence, and the objective function. Based on the total objective function value and the preset total objective function value, determine whether the control algorithm has converged; If the control algorithm fails to converge, the gradient is calculated by backpropagation based on the total objective function value, the predicted control action vector sequence, the second predicted state vector sequence, and the solution function, to obtain the predicted control action vector sequence. Based on the control action amplitude constraint, the predicted control action vector sequence is processed to obtain the updated predicted control action vector sequence; The real-time current state vector, the updated predictive control action vector sequence, and the model identifier are input into the updated edge model. The updated edge model is then called multiple times via recursive calls. The edge model sequentially outputs multiple second predictive state vectors and second predictive system power, forming a second predictive state vector sequence and a second predictive system power sequence; and / or If the control algorithm converges, the predicted control action vector sequence corresponding to the current total objective function value is subjected to safety constraint processing according to the safety constraint function set to obtain the optimal predicted control action vector sequence. The air conditioning operation strategy is executed, which is the first step in executing the optimal predictive control action vector sequence; and / or The objective function is: in, : Overall objective function value; Predict the length of the time series; : The sequence of control actions from time t to t+H-1; Temperature tracking weights; The predicted temperature at time t+k+1; Set the temperature; Energy consumption optimization weights; Predicted system power at time t+k+1; : Control the smoothness weight; : The control action vector at time t+k; and / or The solution function: in, : Overall objective function value; The gradient of the objective function with respect to the control action; The gradient of the objective function with respect to the state; : The Jacobian matrix of state to control action.

10. An air conditioning system, characterized in that, The air conditioning system includes a control module configured to perform the control method as described in any one of claims 1-9.

Images