Energy-saving control method and system for central air-conditioning refrigeration system based on deep learning

Through deep learning technology, the dynamic optimization model is constructed, and the energy efficiency problem of the refrigeration system in the refrigeration room under complex working conditions is solved, the high-efficiency operation of the refrigeration system and sudden abnormal responses are achieved, and the overall energy efficiency and management efficiency are improved.

CN118856530BActive Publication Date: 2024-11-22GUANGZHOU SJ ENERGY SAVING TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411329209.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-24
Publication Date
2024-11-22
Estimated Expiration
2044-09-24

AI Technical Summary

Technical Problem

The existing energy-saving control technology for refrigeration systems in refrigeration rooms is difficult to dynamically adapt to complex and changeable environmental conditions, lacks global optimization capabilities, and cannot effectively coordinate the operation between equipment, resulting in energy waste and inefficient management.

Method used

The energy-saving control method of centralized air-conditioning refrigeration system based on deep learning is adopted, and through technologies such as multi-task learning, adaptive reinforcement learning and federated learning, a dynamic optimization model is built, and the operating parameters of the refrigeration system are monitored and adjusted in real time to achieve global energy efficiency optimization.

Benefits of technology

It realizes the high-efficiency operation of the refrigeration system in the refrigeration room under different loads and environmental conditions, improves the overall energy efficiency of the large-scale distributed refrigeration system, responds to sudden abnormalities in a timely manner, and significantly improves energy-saving effects and management efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118856530B_ABST
    Figure CN118856530B_ABST
Patent Text Reader

Abstract

The present invention proposes an energy-saving control method and system for a centralized air-conditioning refrigeration system based on deep learning, the method comprising: deploying various types of sensors in the refrigeration system of the refrigeration room to monitor key parameters; constructing a deep learning model based on multi-task learning; designing an adaptive reinforcement learning model based on multi-scale state representation and hybrid strategy optimization for optimization; designing a federated learning scheme based on an adaptive weight sharing mechanism to achieve collaborative optimization of the operating parameters of the refrigeration system of the refrigeration room between the refrigeration systems of multiple distributed refrigeration rooms; based on the operation when complex nonlinear behaviors and sudden abnormalities occur in the refrigeration system; defining a multi-objective optimization model to integrate multiple objectives and introduce a risk assessment regularization term to evaluate potential system risks. The present invention provides a new energy-saving control solution by introducing a variety of deep learning and intelligent optimization technologies, which can dynamically adapt to complex and changing environmental conditions and achieve global energy efficiency optimization.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of energy-saving control technology, and in particular to an energy-saving control method and system for a centralized air-conditioning refrigeration system based on deep learning. Background Art

[0002] The refrigeration system of the refrigeration room is one of the indispensable infrastructures in modern industrial and commercial buildings, and is widely used in data centers, cold chain logistics, manufacturing and other fields. These rooms control temperature and humidity through refrigeration equipment to ensure the normal operation of the equipment and the quality of the products. However, the refrigeration system of the refrigeration room is also a high-energy-consuming equipment. During long-term operation, its energy consumption accounts for a large proportion of the total energy consumption of the building. Therefore, optimizing the energy consumption management of the refrigeration system of the refrigeration room is crucial to reducing operating costs and reducing environmental burdens. At present, the energy-saving control methods of the refrigeration system of the refrigeration room mainly rely on preset control strategies and empirical rules. These methods are usually based on historical data or static operating condition models, and achieve a certain degree of energy saving by adjusting the operating parameters of the equipment, such as the start and stop time of the compressor, the refrigerant flow rate, the fan speed, etc.

[0003] However, these traditional control methods have significant limitations. First, the operating conditions and load changes of the refrigeration system in the refrigeration room are complex, and traditional methods are difficult to dynamically adapt to the rapid changes in the environment and load, resulting in unsatisfactory energy-saving effects. Secondly, traditional methods often rely on preset control models or simple feedback control, lack the ability to optimize the system globally, and are prone to fall into local optimality, making it difficult to minimize overall energy consumption. In addition, with the development of big data and intelligent technology, the amount and dimension of data in the refrigeration system of the refrigeration room are increasing. Traditional methods are difficult to fully utilize these data for in-depth analysis and optimization, resulting in energy waste and inefficient management. Finally, traditional control systems are also unable to cope with the complex interactions between devices and cannot effectively coordinate the operation of various devices, further reducing the overall energy efficiency of the system.

[0004] Although some existing improved methods such as model predictive control (MPC) and rule-based adaptive control have improved the flexibility and response speed of control to a certain extent, they still face the following challenges: First, it is difficult to handle the complex nonlinear relationship between multiple variables, resulting in limited system optimization effect; second, in the refrigeration system of multiple machine rooms or large refrigeration rooms, the ability of collaborative control is insufficient, and it is difficult to achieve collaborative optimization in each machine room, resulting in waste of resources; third, for sudden abnormal situations, the traditional system lacks the ability to intelligently identify and respond, and it is difficult to adjust the control strategy in real time to prevent abnormal energy consumption. In summary, there are still many problems that need to be solved in the existing energy-saving control technology of the refrigeration system of the refrigeration room, and there is an urgent need for a more intelligent, efficient and adaptive energy-saving control method for various working conditions. Summary of the invention

[0005] The purpose of this invention is to propose an energy-saving control method and system for a centralized air-conditioning refrigeration system based on deep learning. By introducing a variety of deep learning and intelligent optimization technologies, a new energy-saving control solution is provided, which can dynamically adapt to complex and changeable environmental conditions and achieve global energy efficiency optimization.

[0006] In order to achieve the above object, a first aspect of the present invention provides an energy-saving control method for a centralized air-conditioning refrigeration system based on deep learning, the method comprising:

[0007] S1. Deploy various types of sensors in the refrigeration system of the refrigeration room to monitor key parameters, perform time synchronization processing and preprocessing on the key parameters, and introduce the temperature and humidity combined index and equipment state stability to extract features from the preprocessed data; wherein the key parameters include temperature, humidity, energy consumption and equipment operating status;

[0008] S2. Build a deep learning model based on multi-task learning, divide the deep learning model training task into the task of real-time control and the task of energy consumption prediction and anomaly detection, and add a hybrid regularization strategy to the deep learning model to prevent overfitting of the deep learning model training; the hybrid regularization strategy is combined with Regularization and sparse regularization;

[0009] S3. Design an adaptive reinforcement learning model based on multi-scale state representation and hybrid strategy optimization to optimize the operating parameters of the refrigeration system of the refrigeration room by training with the output of the deep learning model based on multi-task learning;

[0010] S4. Design a federated learning scheme based on an adaptive weight sharing mechanism to achieve collaborative optimization of the refrigeration system operating parameters of the refrigeration room among the refrigeration systems of multiple distributed refrigeration rooms;

[0011] S5. According to the operation when the refrigeration system of the refrigeration room has complex nonlinear behaviors and sudden anomalies, a multi-order autoregressive model based on adaptive time series is constructed according to the loss function of the tasks of energy consumption prediction and anomaly detection to perform energy consumption prediction and anomaly detection on the refrigeration system of the refrigeration room. An adaptive weight adjustment mechanism is used during the training process. If the prediction error of a certain time period is large, the data of the current time period will be given a higher weight in subsequent training. At the same time, a regularization term is designed to improve the stability of the multi-order autoregressive model based on adaptive time series when processing high-dimensional features and reduce the risk of overfitting.

[0012] S6. Define a multi-objective optimization model to integrate multiple objectives and introduce a risk assessment regularization term to evaluate potential system risks. At the same time, set a dynamic weight to adjust the loss function of the multi-objective optimization model to ensure that the refrigeration system of the refrigeration room always focuses on the most important goal under different working conditions. Finally, set feedback control based on reinforcement learning and introduce optimization of the multi-objective optimization model. In addition, introduce a disturbance evaluation mechanism based on Monte Carlo simulation to simulate external disturbances in the refrigeration system of the refrigeration room. Then evaluate and screen the control strategy of the refrigeration system of the refrigeration room to ensure the robustness of the control strategy under various conditions.

[0013] The construction of the adaptive reinforcement learning model is as follows:

[0014] Introduce a multi-scale state representation method to define the state representation, and transform the original state representation Decompose into feature representations at multiple scales , which is expressed as follows:

[0015] ;

[0016] in, represents the feature input at the mth scale, is the feature weight matrix of the current scale;

[0017] Feature representation at multiple scales A hybrid strategy optimization mechanism is designed to handle high-dimensional continuous control problems. The hybrid strategy optimization mechanism is expressed as follows:

[0018] ;

[0019] in, represents the optimal strategy, represents the discount factor, represents the immediate reward, π represents the old strategy, T represents the total time duration, and t represents the current moment;

[0020] A temperature adjustment mechanism is introduced to automatically adjust the exploration rate by monitoring system performance changes and learning rate , which is expressed as follows:

[0021] ;

[0022] in, and denote the minimum and maximum exploration rates, respectively. Indicates adjusting speed parameters.

[0023] Preferably, multiple types of sensors are arranged in different spatial locations and devices to form a multi-dimensional data acquisition matrix , which is expressed as follows:

[0024] ;

[0025] in, represents the temperature at position i, represents the humidity of the i-th device, represents the power consumption of the i-th device, Indicates the status of the i-th compressor, 0 means off, 1 means on, represents the fan speed of the i-th device, and n is the total number of sensors;

[0026] The time synchronization process is represented as follows:

[0027] ;

[0028] in, represents the data of the i-th sensor processed in time synchronization, represents the raw data of the i-th sensor, is the time delay of the sensor, is the synchronization adjustment coefficient, It is a dynamic time offset item calculated based on historical data and is used to refine time synchronization;

[0029] The preprocessing includes data calibration and noise processing, data cleaning and missing value filling and data standardization.

[0030] Preferably, the temperature and humidity combined index Combining the characteristics of temperature and humidity, it is expressed as follows:

[0031] ;

[0032] The stability of the equipment state Reflects the stability of the equipment's operating status, as shown below:

[0033] ;

[0034] Where T represents the total number of windows and k represents the kth data point.

[0035] Preferably, the loss function of the deep learning model based on multi-task learning is expressed as follows:

[0036] ;

[0037] in, represents the loss function of the control task, represents the loss function of energy consumption prediction and anomaly detection tasks, represents the regularization term, which is used to prevent overfitting. , and represents the weight coefficient between tasks, Represents the model parameters of a deep learning model based on multi-task learning;

[0038] Among them, the loss function of the control task is , which is expressed as follows:

[0039] ;

[0040] in, represents the immediate reward obtained at time step t, represents the discount factor at time step t, which is used to balance short-term and long-term benefits;

[0041] The loss function for the energy consumption prediction and anomaly detection tasks is , which is expressed as follows:

[0042] ;

[0043] in, represents the energy consumption value predicted by the model, Indicates the actual energy consumption value, represents the regularization coefficient, Represents the second-order derivative of the predicted output with respect to the input feature, M represents the total number of features, and N represents the total number of samples;

[0044] The hybrid regularization strategy is expressed as follows:

[0045] ;

[0046] in, represents the hybrid regularization term, express Regularization term, represents the kth feature weight Regularization term, K represents the number of features, and represents the corresponding regularization coefficient.

[0047] Preferably, the S4 specifically includes:

[0048] S401, design the training objectives of the local models in each computer room, where the goal of the local model in each computer room is to minimize the local loss function , which is expressed as follows:

[0049] ;

[0050] in, and represents the data and label of the i-th training sample, represents the model parameters of computer room k, represents the regularization coefficient, Represents the personalized feature matrix of the computer room, represents element-wise product, represents the loss function, Indicates the actual operation data of the computer room. Indicates the operating characteristics of the computer room;

[0051] S402: After each computer room completes local model training, an adaptive weight sharing mechanism is used to perform collaborative optimization on a global scale, as shown below:

[0052] ;

[0053] in, represents the global model parameters, represents the adaptive weight related to the data quality and model performance of computer room k;

[0054] S403: Update the global model parameters The feedback mechanism is used to transmit the information back to each computer room. At the same time, a momentum correction mechanism is designed to adjust the degree of integration of new and old parameters. As the initial parameters for the next round of local training; wherein the momentum correction mechanism is expressed as follows:

[0055] ;

[0056] in, Represents the momentum factor, which controls the weight ratio of new and old parameters.

[0057] Preferably, the adaptive weight , calculated as follows:

[0058] ;

[0059] in, represents the adjustment parameter that controls the sensitivity of the weight distribution, represents the global model parameters of computer room k, represents the global model parameters of computer room j, represents the local loss function of computer room k, Represents the local loss function of computer room j.

[0060] Preferably, the prediction formula of the multi-order autoregressive model based on adaptive time series is expressed as follows:

[0061] ;

[0062] in, represents the predicted energy consumption value at time step t, represents the actual energy consumption value of the previous i time steps, represents the autoregressive coefficient, represents the moving average coefficient, represents the residual of the first j time steps, represents the offset term; Represents a multi-dimensional feature correction term, where represents the multidimensional feature vector at time step t, Represents a multi-dimensional feature correction item The corresponding weight vector;

[0063] A multi-task loss function based on a multi-order autoregressive model of adaptive time series is designed. The multi-task loss function includes energy consumption prediction loss and anomaly detection loss, which is expressed as follows:

[0064] ;

[0065] in, and Represent the weight coefficients of energy consumption prediction error and anomaly detection loss, respectively. represents the threshold used to detect anomalies, represents the standard deviation of the residuals, which is used to weight the confidence of outliers. represents the actual energy consumption value at time step t;

[0066] in, To minimize the prediction error, By adjusting the confidence, higher detection weights are given to abnormal situations.

[0067] Preferably, in the training process of the multi-order autoregressive model based on adaptive time series, if the prediction error of a certain time period is large, the data of the current time period will be given the highest weight in subsequent training, wherein the weight adjustment formula is expressed as follows:

[0068] ;

[0069] in, represents the sample weight at time step t, represents the sensitivity parameter for weight adjustment;

[0070] The design regularization term improves the stability of the multi-order autoregressive model based on adaptive time series when processing high-dimensional features. , which is expressed as follows:

[0071] ;

[0072] in, represents the weight vector of the i-th feature, , and They represent the regularization hyperparameters, Represents a small constant used to prevent the denominator from being zero.

[0073] Preferably, the comprehensive loss function of the multi-objective optimization model is It is expressed as follows:

[0074] ;

[0075] in, represents the energy consumption at time step t, Indicates the cooling efficiency, represents the standard deviation of temperature fluctuations, represents the rate of change of temperature, represents the risk assessment regularization term, which is used to evaluate potential system risks. , , , Respectively represent the weight coefficient of each target; wherein the targets include energy consumption, cooling efficiency and system stability;

[0076] The risk assessment regularization term It is expressed as follows:

[0077] ;

[0078] in, represents the kth control parameter of the system, represents the risk weight coefficient, represents the second-order derivative of energy consumption to the control parameter, which is used to capture the sensitivity of the system. Indicates the sensitivity of temperature fluctuations to control parameters;

[0079] The dynamic weight adjustment multi-objective optimization model loss function is set to automatically adjust the weight coefficient of each objective in the loss function according to the real-time status of the system in actual operation. , which is expressed as follows:

[0080] ;

[0081] in, represents the initial base weight, represents the adjustment coefficient, Represents the disturbance term at time step t, which is used to introduce a certain amount of randomness to prevent falling into the local optimal solution.

[0082] In a second aspect of the present invention, a deep learning-based refrigeration system energy-saving control system for a refrigeration room is provided, the system comprising:

[0083] The parameter collection module is used to deploy various types of sensors in the refrigeration system of the refrigeration room to monitor key parameters, perform time synchronization processing and preprocessing on the key parameters, and introduce the temperature and humidity combined index and equipment state stability to extract features from the preprocessed data; wherein the key parameters include temperature, humidity, energy consumption and equipment operating status;

[0084] The energy consumption prediction and anomaly detection module is used to build a deep learning model based on multi-task learning, which divides the deep learning model training task into the task of real-time control and the task of energy consumption prediction and anomaly detection. At the same time, a hybrid regularization strategy is added to the deep learning model to prevent overfitting of the deep learning model training; the hybrid regularization strategy is combined with Regularization and sparse regularization;

[0085] An operating parameter optimization module is used to design an adaptive reinforcement learning model based on multi-scale state representation and hybrid strategy optimization to optimize the operating parameters of the refrigeration system of the refrigeration room by training with the output of a deep learning model based on multi-task learning;

[0086] A federated model building module is used to design a federated learning scheme based on an adaptive weight sharing mechanism to achieve collaborative optimization of the operating parameters of the refrigeration system of multiple distributed refrigeration rooms;

[0087] The sudden anomaly optimization module is used to perform energy consumption prediction and anomaly detection on the refrigeration system of the refrigeration room according to the loss function of the energy consumption prediction and anomaly detection tasks when the refrigeration system of the refrigeration room has complex nonlinear behaviors and sudden anomalies. The module also uses an adaptive time series multi-order autoregressive model to perform energy consumption prediction and anomaly detection on the refrigeration system of the refrigeration room. In the training process, an adaptive weight adjustment mechanism is used. If the prediction error of a certain time period is large, the data of the current time period will be given a higher weight in subsequent training. At the same time, a regularization term is designed to improve the stability of the multi-order autoregressive model based on adaptive time series when processing high-dimensional features and reduce the risk of overfitting.

[0088] The system optimization module is used to define a multi-objective optimization model to integrate multiple objectives and introduce a risk assessment regularization term to evaluate potential system risks. At the same time, dynamic weights are set to adjust the loss function of the multi-objective optimization model to ensure that the refrigeration system of the refrigeration room always focuses on the most important goal under different working conditions. Finally, feedback control based on reinforcement learning is set and the multi-objective optimization model is optimized. A disturbance evaluation mechanism based on Monte Carlo simulation is introduced to simulate external disturbances in the refrigeration system of the refrigeration room. Then, the control strategy of the refrigeration system of the refrigeration room is evaluated and screened to ensure the robustness of the control strategy under various conditions.

[0089] The construction of the adaptive reinforcement learning model is as follows:

[0090] Introduce a multi-scale state representation method to define the state representation, and transform the original state representation Decompose into feature representations at multiple scales , which is expressed as follows:

[0091] ;

[0092] in, represents the feature input at the mth scale, is the feature weight matrix of the current scale;

[0093] Feature representation at multiple scales A hybrid strategy optimization mechanism is designed to handle high-dimensional continuous control problems. The hybrid strategy optimization mechanism is expressed as follows:

[0094] ;

[0095] in, represents the optimal strategy, represents the discount factor, represents the immediate reward, π represents the old strategy, T represents the total time duration, and t represents the current moment;

[0096] A temperature adjustment mechanism is introduced to automatically adjust the exploration rate by monitoring system performance changes and learning rate , which is expressed as follows:

[0097] ;

[0098] in, and denote the minimum and maximum exploration rates, respectively. Indicates adjusting speed parameters.

[0099] The beneficial technical effects of the present invention are at least as follows:

[0100] The present invention uses an adaptive reinforcement learning model to optimize the control strategy of the refrigeration system in the refrigeration room in real time, solving the problem that traditional methods are difficult to dynamically respond to changes in complex working conditions. The current model continuously learns and adjusts control parameters such as compressor start time, refrigerant flow, fan speed, etc., to ensure that the system always maintains high-efficiency operation under different load and environmental conditions, overcoming the limitations of static control of traditional methods.

[0101] The present invention adopts federated learning technology to achieve collaborative optimization control of multiple computer rooms. Each computer room shares some parameters through a locally trained deep learning model, and jointly optimizes the overall energy-saving strategy while ensuring data privacy, solving the problem of insufficient collaborative control capabilities of multiple computer rooms in the prior art, and significantly improving the overall energy efficiency of the refrigeration system of large-scale distributed refrigeration computer rooms.

[0102] The present invention also introduces a generative adversarial network (GAN) for energy consumption prediction and anomaly detection. Through the generator and discriminator of the GAN model, the system can accurately predict future energy consumption trends, monitor energy consumption anomalies in real time, provide early warnings and automatic adjustment strategies, and solve the problem of traditional systems not responding in time and adjusting unintelligently when dealing with sudden anomalies.

[0103] The present invention uses deep transfer learning technology to apply models trained in other similar scenarios to new computer rooms, reducing the time and cost of model retraining while ensuring efficient control effects. Transfer learning effectively solves the problem of poor model generalization ability of traditional systems in changing environments, allowing the system to quickly adapt to different application scenarios.

[0104] The present invention provides a highly intelligent energy-saving control system for the refrigeration system of the refrigeration room by integrating the latest technologies such as adaptive reinforcement learning, federated learning, generative adversarial networks, and deep transfer learning. The current system can dynamically adapt to changing working conditions, collaboratively optimize the energy-saving strategies of multiple rooms, and quickly respond and adjust the control strategy in the event of sudden abnormalities, achieving the optimal configuration of global energy efficiency and greatly improving the energy-saving effect and management efficiency of the refrigeration system of the refrigeration room. BRIEF DESCRIPTION OF THE DRAWINGS

[0105] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0106] Figure 1 This is a flow chart of the energy-saving control method of a centralized air-conditioning refrigeration system based on deep learning of the present invention.

[0107] Figure 2This is a framework diagram of the energy-saving control system of the refrigeration system of the refrigeration room based on deep learning of the present invention. DETAILED DESCRIPTION

[0108] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0109] like Figure 1 As shown, an energy-saving control method for a centralized air-conditioning refrigeration system based on deep learning is provided in an embodiment of the present invention, and the method includes the following steps S1-S6.

[0110] S1. Various types of sensors are deployed in the refrigeration system of the refrigeration room to monitor key parameters, and the key parameters are time-synchronized and pre-processed. The temperature and humidity combined index and equipment state stability are introduced to extract features from the pre-processed data; wherein the key parameters include temperature, humidity, energy consumption and equipment operating status.

[0111] Specifically, various types of sensors are deployed in the refrigeration system of the refrigeration room to monitor key parameters, including temperature, humidity, energy consumption, and equipment operating status. These sensors are arranged in different spatial locations and devices, forming a multi-dimensional data collection matrix , whose elements are represented by the following formula:

[0112] ;

[0113] in, represents the temperature at position i, Indicates humidity, represents the power consumption of the i-th device, Indicates the status of the compressor (0 means off, 1 means on), Indicates the fan speed, and n is the total number of sensors.

[0114] Among them, in multi-dimensional sensor data collection, the data collection time of different sensors may be out of sync, resulting in inconsistency of time series data. To this end, the present invention introduces an innovative time synchronization mechanism to synchronize data based on the following formula:

[0115] ;

[0116] in, represents the raw data of the i-th sensor, is the time delay of the sensor, is the synchronization adjustment coefficient, It is a dynamic time offset item calculated based on historical data and is used to refine time synchronization.

[0117] Furthermore, in the actual scenario of the refrigeration system in the refrigeration room, the sensor data may contain noise and errors due to factors such as equipment vibration, electromagnetic interference, etc. To this end, the present invention proposes an innovative calibration and noise processing method, which combines regularization terms to improve the stability and reliability of the data.

[0118] Among them, the sensor is calibrated by the following formula, using a double adjustment mechanism:

[0119] ;

[0120] in, and is the calibration coefficient of the sensor, It is a calibration adjustment factor based on energy consumption data, which is used to correct the deviation of temperature and humidity data caused by changes in the operating status of the equipment. This dual adjustment mechanism can better adapt to the changes in sensor data when the equipment is frequently started and stopped.

[0121] Furthermore, in order to solve the noise problem, the present invention adopts a noise suppression method based on regularization. First, the signal after noise filtering is calculated using the following formula:

[0122] ;

[0123] in, is a regularization parameter that controls the smoothness of the data. The current regularization term is mainly used to reduce high-frequency noise in the data while maintaining the trend characteristics of the signal.

[0124] Furthermore, the data of the refrigeration system of the refrigeration room has missing and abnormal situations caused by sensor failure or temporary failure. To this end, the present invention designs an advanced data cleaning and filling strategy, which realizes high-precision filling of missing values ​​through joint modeling of historical trends and local features.

[0125] The detection of abnormal data is based on a statistical model. The present invention uses the following formula to determine whether the data is abnormal:

[0126] ;

[0127] in, for The moving average value within the window period, is the corresponding standard deviation, is the threshold coefficient for anomaly detection. If established, are marked as outliers and removed.

[0128] For missing values, the present invention proposes a filling strategy based on local weighted linear interpolation. The specific formula is as follows:

[0129] ;

[0130] Among them, K is the number of historical data points used for filling, is the weighting coefficient, Control the decay speed. This method makes full use of historical data close in time and improves the accuracy of filling.

[0131] Furthermore, in order to improve the training effect of subsequent deep learning models, the present invention designs an innovative feature extraction scheme based on data standardization, and proposes two new features based on the characteristics of the refrigeration system of the refrigeration room.

[0132] The formula for data normalization is: ;

[0133] in, and They are The mean and standard deviation in the historical data.

[0134] Furthermore, based on the traditional moving average and fluctuation amplitude, the present invention introduces two new features: temperature and humidity combined index (THI) and equipment state stability (ESI).

[0135] Among them, the temperature and humidity combined index (THI) combines the characteristics of temperature and humidity and is calculated as follows:

[0136] ;

[0137] The current index can comprehensively reflect the thermal comfort of the environment and is suitable for evaluating the effectiveness of the refrigeration system in the refrigeration room.

[0138] Equipment status stability index (ESI) reflects the stability of the equipment operation status, and the formula is:

[0139] ;

[0140] The current index shows the relationship between the frequency of device start-up and shutdown and power consumption within a time window, and can identify the impact of high-frequency start-up and shutdown on energy consumption.

[0141] S2. Build a deep learning model based on multi-task learning, divide the deep learning model training task into the task of real-time control and the task of energy consumption prediction and anomaly detection, and add a hybrid regularization strategy to the deep learning model to prevent overfitting of the deep learning model training; the hybrid regularization strategy is combined with Regularization and sparse regularization, including the following steps:

[0142] First, the loss function of multi-task learning is defined to jointly optimize different tasks. The present invention divides the model into two parts: one is for the task of real-time control, and the other is for the task of energy consumption prediction and anomaly detection. The loss function is designed as:

[0143] ;

[0144] in, is the loss function of the control task, is the loss function for energy consumption prediction and anomaly detection tasks, is a regularization term used to prevent overfitting. , and is the weight coefficient between tasks, are model parameters.

[0145] Specifically, the loss function of the control task The dynamic control strategy for optimizing the system is defined by the following formula using a reinforcement learning-based framework:

[0146] ;

[0147] in, is the immediate reward obtained at time step t, is the discount factor used to balance short-term and long-term benefits.

[0148] This loss function guides the model to learn to adjust the parameters of the refrigeration system in the refrigeration room during training to achieve the optimal balance between energy consumption and performance.

[0149] In energy consumption prediction and anomaly detection tasks, the loss function Designed for:

[0150] ;

[0151] in, is the energy consumption value predicted by the model, is the actual energy consumption value, is the regularization coefficient, is the second-order derivative of the predicted output with respect to the input feature. This regularization term can suppress the model’s oversensitivity to high-frequency changes and enhance the model’s robustness in anomaly detection.

[0152] Finally, the regularization term It is used to control model complexity, prevent overfitting, and guide the model to learn a smoother and more consistent control strategy. This paper introduces an innovative hybrid regularization strategy, combining Regularization and Sparse Regularization:

[0153] ;

[0154] in, yes Regularization term, is the kth feature weight Regularization term, K is the number of features, and is the corresponding regularization coefficient. This hybrid regularization strategy can balance the complexity and sparsity of the model, ensuring that the model can fully learn complex nonlinear relationships while avoiding overfitting.

[0155] S3. Design an adaptive reinforcement learning model based on multi-scale state representation and hybrid strategy optimization to optimize the operating parameters of the refrigeration system of the refrigeration room by training with the output of the deep learning model based on multi-task learning.

[0156] Specifically, the control parameters of the system are dynamically optimized to adapt to environmental changes in real time. After completing the training of the deep learning model, the adaptive reinforcement learning control stage is entered. The goal of this step is to dynamically optimize the operating parameters of the refrigeration system of the refrigeration room through the trained deep learning model to achieve real-time energy saving and efficient control.

[0157] The following steps are involved:

[0158] First, define the state representation. In the refrigeration system of the refrigeration room, the state space It usually has high dimensions and contains multi-scale information, such as temperature, humidity, energy consumption and device status. Therefore, the present invention introduces a multi-scale state representation method to represent the original state. Decompose into feature representations at multiple scales :

[0159] ;

[0160] in, represents the feature input at the mth scale, is the feature weight matrix of the current scale. Through this multi-scale representation, the model can capture the dynamic characteristics of the system at different time scales.

[0161] Then, based on the state representation, the present invention designs a hybrid policy optimization mechanism. The current mechanism combines the policy optimization based on value function and the direct policy optimization based on policy gradient to better handle high-dimensional continuous control problems. The core formula of policy optimization is:

[0162] ;

[0163] in, represents the optimal strategy, is the discount factor, This invention combines the deep Q network (DQN) and the policy gradient method to gradually approach the optimal strategy using a hybrid strategy optimization mechanism. , in order to achieve dynamic optimization of control strategy.

[0164] Finally, in order to enhance the robustness and adaptability of the model, this paper introduces an innovative temperature adjustment mechanism to dynamically adjust the learning rate and exploration strategy. The current mechanism automatically adjusts the exploration rate by monitoring the performance changes of the system. and learning rate , the announcement is as follows:

[0165] ;

[0166] in, and are the minimum and maximum exploration rates, is the speed parameter. The current mechanism ensures sufficient exploration in the early stages and convergence to the optimal strategy in the later stages.

[0167] S4. Design a federated learning scheme based on an adaptive weight sharing mechanism to achieve collaborative optimization of the operating parameters of the refrigeration system of a refrigeration room between the refrigeration systems of multiple distributed refrigeration rooms. Collaborative optimization cannot be achieved between multiple computer rooms, resulting in low overall energy efficiency. After completing the design of the adaptive reinforcement learning control, enter the stage of collaborative optimization of multi-computer room federated learning. The purpose of this step is to achieve collaborative optimization between the refrigeration systems of multiple distributed refrigeration rooms to further improve the energy efficiency of the overall system. To this end, the present invention designs a federated learning scheme based on an adaptive weight sharing mechanism to ensure that the personalized needs of each computer room and the global optimization goals can be effectively met.

[0168] The following steps are involved:

[0169] First, define the training objectives of the local models in each computer room. In each computer room k, the model’s goal is to minimize the local loss function The current function combines the actual operation data of the computer room and its unique operating characteristics , calculated by the following formula:

[0170] ;

[0171] in, and is the data and label of the i-th training sample, is the model parameter of computer room k, is the regularization coefficient, It is the personalized feature matrix of the computer room. represents element-wise product, Represents the loss function. The regular term in this formula ensures the personalized adaptability of the model.

[0172] After each computer room completes local model training, the present invention uses an adaptive weight sharing mechanism to perform collaborative optimization on a global scale. Global model parameters Updated by the following formula:

[0173] ;

[0174] in, is an adaptive weight related to the data quality and model performance of computer room k, and is calculated as follows:

[0175] ;

[0176] in, is a tuning parameter that controls the sensitivity of weight distribution. Through this adaptive weight distribution mechanism, models with better performance contribute more to the global update, thereby improving the overall optimization effect while ensuring personalization.

[0177] Finally, the updated global model parameters The feedback mechanism is used to transmit the initial parameters to each computer room for the next round of local training. In order to ensure smooth transition and stable convergence, the present invention designs a momentum correction mechanism to gradually adjust the degree of integration of new and old parameters. The formula is as follows:

[0178] ;

[0179] in, It is the momentum factor, which controls the weight ratio of new and old parameters to ensure the smoothness of the update process and the stability of the system.

[0180] S5. According to the operation of the refrigeration system of the refrigeration room when complex nonlinear behaviors and sudden anomalies occur, a multi-order autoregressive model based on adaptive time series is constructed according to the loss function of the tasks of energy consumption prediction and anomaly detection to perform energy consumption prediction and anomaly detection on the refrigeration system of the refrigeration room. An adaptive weight adjustment mechanism is used during the training process. If the prediction error of a certain time period is large, the data of the current time period will be given a higher weight in subsequent training. At the same time, a regularization term is designed to improve the stability of the multi-order autoregressive model based on adaptive time series when processing high-dimensional features and reduce the risk of overfitting.

[0181] Specifically, in order to deal with the complex nonlinear behaviors and sudden anomalies that may occur in the refrigeration system of the refrigeration room, a new solution was designed, including the following steps:

[0182] In the actual application of the refrigeration system in the refrigeration room, the energy consumption data usually has a high time series correlation and may be affected by multiple factors (such as ambient temperature, equipment operating status, etc.). Therefore, the present invention proposes a multi-order autoregressive model, and on this basis, adds an adaptive correction term of multi-dimensional feature fusion. The prediction formula is:

[0183] ;

[0184] in, is the predicted energy consumption at time step t, is the actual energy consumption value of the previous i time steps, is the autoregressive coefficient, is the moving average coefficient, is the residual of the first j time steps, is the offset term. Innovatively introduces multi-dimensional feature correction terms ,in is a multidimensional feature vector at time step t (such as temperature, humidity, equipment load, etc.), is the corresponding weight vector. This correction can capture the impact of complex environmental variables on energy consumption and achieve more accurate prediction.

[0185] In order to ensure that the model can not only accurately predict conventional energy consumption but also detect anomalies in a timely manner, the present invention designs an innovative multi-task loss function in the prediction model, aiming to simultaneously optimize the energy consumption prediction accuracy and the sensitivity of anomaly detection.

[0186] Furthermore, the multi-task loss function proposed in the present invention consists of two parts: energy consumption prediction loss and anomaly detection loss. The core formula of the loss function is:

[0187] ;

[0188] in, and are the weight coefficients of energy consumption prediction error and anomaly detection loss, respectively. is the threshold used to detect anomalies, is the standard deviation of the residuals, which is used to weight the confidence of the outliers. The first loss term is used to minimize the prediction error, and the second loss term is adjusted by the confidence to give higher detection weights to possible anomalies. This design ensures that the model can be sensitive to and respond to anomalies (such as equipment failures or emergencies) while maintaining high prediction accuracy.

[0189] Furthermore, in view of the uncertainty in the operation of the refrigeration system in the refrigeration room, an adaptive weight adjustment mechanism is introduced to improve the robustness and generalization ability of the model in different time periods. This mechanism is particularly suitable for complex situations such as periodic fluctuations and sudden failures that may occur in the refrigeration system of the refrigeration room. During the training process, if the prediction error of a certain time period is large, the data of the current time period will be given a higher weight in subsequent training to improve the model's adaptability to difficult samples. The weight adjustment formula is:

[0190] ;

[0191] in, is the sample weight at time step t, is the sensitivity parameter for weight adjustment. The current formula amplifies the weight of time periods with large prediction errors in training, ensuring that the model gradually enhances its prediction ability for these time periods. This mechanism can dynamically balance the model's emphasis on each time period and improve the overall prediction and anomaly detection performance.

[0192] Furthermore, the present invention introduces an innovative regularization term into the model to further improve the stability of the model when processing high-dimensional features and reduce the risk of overfitting. This regularization term combines sparsity and complexity constraints, and the designed regularization term is as follows:

[0193] ;

[0194] in, is the weight vector of the ith feature, , and is a regularization hyperparameter, is a small constant used to prevent the denominator from being zero. The current regularization term combines the sparsity ( Regularization), complexity constraints ( Regularization) and adaptive sparsity adjustment ensure that the model does not overfit while maintaining high predictive ability, which is particularly suitable for energy consumption prediction in high-dimensional feature environments.

[0195] The innovative solution of this step not only achieves high-precision energy consumption prediction, but also enhances the detection capability of abnormal energy consumption events through multi-order autoregressive model combined with multi-dimensional feature correction, adaptive weight adjustment mechanism and innovative regularization term design. This solution further improves the intelligence level of the refrigeration system of the refrigeration room on the basis of the previous steps, and provides strong support for the energy-saving control goal of the entire patent solution.

[0196] S6. Define a multi-objective optimization model to integrate multiple objectives and introduce a risk assessment regularization term to evaluate potential system risks. At the same time, set dynamic weights to adjust the loss function of the multi-objective optimization model to ensure that the refrigeration system of the refrigeration room always focuses on the most important goal under different working conditions. Finally, set feedback control based on reinforcement learning and introduce optimization of the multi-objective optimization model. In addition, introduce a disturbance evaluation mechanism based on Monte Carlo simulation to simulate external disturbances in the refrigeration system of the refrigeration room. Then, evaluate and screen the control strategy of the refrigeration system of the refrigeration room to ensure the robustness of the control strategy under various conditions.

[0197] Specifically, the following steps are included:

[0198] First, define a comprehensive loss function , which integrates multiple objectives such as energy consumption, refrigeration efficiency, and system stability. This loss function will serve as the core of the multi-objective optimization algorithm of the present invention to help the system achieve the optimal control strategy under different working conditions. The formula is as follows:

[0199] ;

[0200] in, is the energy consumption at time step t, is the cooling efficiency, is the standard deviation of temperature fluctuations, is the rate of change of temperature, is an innovative risk assessment regularization term used to assess potential system risks. , , , Represent the weight coefficients of each objective. This formula not only balances energy consumption and efficiency, but also System security considerations are introduced.

[0201] Current regularization term The design of takes into account a variety of potential risk factors, and the formula is as follows:

[0202] ;

[0203] in, is the kth control parameter of the system, is the risk weight coefficient, represents the second-order derivative of energy consumption to the control parameter, which is used to capture the sensitivity of the system. Represents the sensitivity of temperature fluctuations to control parameters. This regularization term ensures that during the optimization process, the system can effectively respond to potential risks and avoid runaway or unstable situations under extreme conditions.

[0204] Next, a multi-objective optimization algorithm based on dynamic weight adjustment is used to automatically adjust the weight coefficients of each objective in the loss function according to the real-time status of the system in actual operation. The current adjustment mechanism can ensure that the system always focuses on the most important goal under different working conditions. The specific formula is:

[0205] ;

[0206] in, is the initial base weight, is the adjustment coefficient, is the disturbance term at time step t, which is used to introduce a certain amount of randomness to prevent falling into the local optimal solution. Through this dynamic adjustment mechanism, the system can automatically adjust the strategy in different environments and better cope with complex and changing operating conditions.

[0207] In order to enhance the real-time response capability and adaptability of the system, this paper designs a feedback control mechanism based on reinforcement learning. The current mechanism combines deep Q learning (DQN) with the policy gradient method to dynamically adjust the control strategy by real-time monitoring of the key indicators of the system. The specific adjustment strategy formula is:

[0208] ;

[0209] in, is the current control parameter vector, represents the gradient of the control parameter, is the learning rate, is the current state vector of the system. Through this feedback control mechanism, the system can automatically correct the control strategy according to the derivative information of the comprehensive loss function at each time step, thereby ensuring the optimal operating state of the system in a dynamic environment.

[0210] Finally, in order to ensure the stability of the system in complex environments, a disturbance assessment mechanism based on Monte Carlo simulation is also introduced. The current mechanism evaluates and screens the control strategy of the system by simulating possible external disturbances (such as load fluctuations, sudden changes in ambient temperature, etc.) to ensure its robustness under various conditions. The disturbance assessment formula is:

[0211] ;

[0212] in, is an assessment of the stability of the current strategy. represents external disturbance, with mean 0 and variance Through this mechanism, the system can screen out the best strategy under different disturbance conditions, thereby improving the robustness in actual operation.

[0213] Through this comprehensive intelligent decision-making and feedback control solution, not only the global optimization control of the refrigeration system of the refrigeration room is achieved, but also the efficient and stable operation of the system in a complex environment is ensured. The current solution is closely linked to the first five steps: through the early warning information provided by energy consumption prediction and anomaly detection, the control model optimized by federated learning, and the dynamic strategy generated by adaptive reinforcement learning, the comprehensive intelligent decision-making system can effectively integrate the advantages of each part to achieve global optimal control. In addition, the regularization term and The introduction of has increased the ability of risk control and stability assessment for the system, making up for the shortcomings of traditional control strategies in extreme situations. The design of these innovative formulas makes the entire patented solution more flexible and intelligent in dealing with actual working conditions. In summary, the comprehensive intelligent decision-making and feedback control solution combines multi-objective dynamic optimization, adaptive feedback control and Monte Carlo simulation to ensure the global optimal operation state of the system in a complex environment.

[0214] To better implement the above method, Figure 2 As shown, the present invention also proposes a refrigeration system energy-saving control system for a refrigeration room based on deep learning, and the system includes:

[0215] The parameter collection module 101 is used to deploy various types of sensors in the refrigeration system of the refrigeration room to monitor key parameters, perform time synchronization processing and preprocessing on the key parameters, and introduce the temperature and humidity combined index and the equipment state stability to extract features from the preprocessed data; wherein the key parameters include temperature, humidity, energy consumption and the operating state of the equipment;

[0216] The energy consumption prediction and anomaly detection module 102 is used to build a deep learning model based on multi-task learning, divide the deep learning model training task into the task of real-time control and the task of energy consumption prediction and anomaly detection, and add a hybrid regularization strategy to the deep learning model to prevent overfitting of the deep learning model training; the hybrid regularization strategy is combined with Regularization and sparse regularization;

[0217] An operating parameter optimization module 103 is used to design an adaptive reinforcement learning model based on multi-scale state representation and hybrid strategy optimization to optimize the operating parameters of the refrigeration system of the refrigeration room by training with the output of the deep learning model based on multi-task learning;

[0218] A federated model building module 104 is used to design a federated learning scheme based on an adaptive weight sharing mechanism to achieve collaborative optimization of refrigeration system operating parameters of refrigeration rooms among refrigeration systems of multiple distributed refrigeration rooms;

[0219] The sudden abnormality optimization module 105 is used to construct a multi-order autoregressive model based on adaptive time series according to the loss function of the energy consumption prediction and anomaly detection tasks to perform energy consumption prediction and anomaly detection on the refrigeration system of the refrigeration room according to the operation when the refrigeration system of the refrigeration room has complex nonlinear behaviors and sudden abnormalities. In the training process, an adaptive weight adjustment mechanism is used. If the prediction error of a certain time period is large, the data of the current time period will be given a higher weight in subsequent training. At the same time, a regularization term is designed to improve the stability of the multi-order autoregressive model based on adaptive time series when processing high-dimensional features and reduce the risk of overfitting;

[0220] The system optimization module 106 is used to define a multi-objective optimization model to integrate multiple objectives and introduce a risk assessment regularization term to evaluate potential system risks. At the same time, a dynamic weight is set to adjust the loss function of the multi-objective optimization model to ensure that the refrigeration system of the refrigeration room always focuses on the most important goal under different working conditions. Finally, a feedback control based on reinforcement learning is set and the multi-objective optimization model is optimized. A disturbance evaluation mechanism based on Monte Carlo simulation is introduced to simulate the external disturbance of the refrigeration system of the refrigeration room. Then, the control strategy of the refrigeration system of the refrigeration room is evaluated and screened to ensure the robustness of the control strategy under various conditions.

[0221] The construction of the adaptive reinforcement learning model is as follows:

[0222] Introduce a multi-scale state representation method to define the state representation, and transform the original state representation Decompose into feature representations at multiple scales , which is expressed as follows:

[0223] ;

[0224] in, represents the feature input at the mth scale, is the feature weight matrix of the current scale;

[0225] Feature representation at multiple scales A hybrid strategy optimization mechanism is designed to handle high-dimensional continuous control problems. The hybrid strategy optimization mechanism is expressed as follows:

[0226] ;

[0227] in, represents the optimal strategy, represents the discount factor, represents the immediate reward, π represents the old strategy, T represents the total time duration, and t represents the current moment;

[0228] A temperature adjustment mechanism is introduced to automatically adjust the exploration rate by monitoring system performance changes and learning rate , which is expressed as follows:

[0229] ;

[0230] in, and denote the minimum and maximum exploration rates, respectively. Indicates adjusting speed parameters.

[0231] Unless otherwise specifically stated, the relative steps, numerical expressions and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present application.

[0232] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present application can essentially be embodied in the form of a software product, or in other words, the part that contributes to the prior art or the part of the current technical solution. The current computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, and other media that can store program codes.

[0233] In the description of the present application, it should be noted that the terms "upper", "lower", etc. indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, or are orientations or positional relationships that are commonly placed when the current invention products are used. They are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0234] In the description of this application, it should also be noted that, unless otherwise clearly specified and limited, the terms "set", "install", "connect", and "connect" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two elements. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0235] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A deep learning-based energy-saving control method for a centralized air-conditioning refrigeration system, characterized in that: The method comprises: S1. Deploy various types of sensors in the refrigeration system of the refrigeration room to monitor key parameters, perform time synchronization processing and preprocessing on the key parameters, and introduce the temperature and humidity combined index and equipment state stability to extract features from the preprocessed data; wherein the key parameters include temperature, humidity, energy consumption and equipment operating status; S2. Build a deep learning model based on multi-task learning, divide the deep learning model training task into the task of real-time control and the task of energy consumption prediction and anomaly detection, and add a hybrid regularization strategy to the deep learning model to prevent overfitting of the deep learning model training; the hybrid regularization strategy is combined with Regularization and sparse regularization; S3. Design an adaptive reinforcement learning model based on multi-scale state representation and hybrid strategy optimization to optimize the operating parameters of the refrigeration system of the refrigeration room by training with the output of the deep learning model based on multi-task learning; S4. Design a federated learning scheme based on an adaptive weight sharing mechanism to achieve collaborative optimization of the refrigeration system operating parameters of the refrigeration room among the refrigeration systems of multiple distributed refrigeration rooms; S5. According to the operation when the refrigeration system of the refrigeration room has complex nonlinear behaviors and sudden anomalies, a multi-order autoregressive model based on adaptive time series is constructed according to the loss function of the tasks of energy consumption prediction and anomaly detection to perform energy consumption prediction and anomaly detection on the refrigeration system of the refrigeration room. An adaptive weight adjustment mechanism is used during the training process. If the prediction error of a certain time period is large, the data of the current time period will be given a higher weight in subsequent training. At the same time, a regularization term is designed to improve the stability of the multi-order autoregressive model based on adaptive time series when processing high-dimensional features and reduce the risk of overfitting. S6. Define a multi-objective optimization model to integrate multiple objectives and introduce a risk assessment regularization term to evaluate potential system risks. At the same time, set a dynamic weight to adjust the loss function of the multi-objective optimization model to ensure that the refrigeration system of the refrigeration room always focuses on the most important goal under different working conditions. Finally, set feedback control based on reinforcement learning and introduce optimization of the multi-objective optimization model. In addition, introduce a disturbance evaluation mechanism based on Monte Carlo simulation to simulate external disturbances in the refrigeration system of the refrigeration room. Then evaluate and screen the control strategy of the refrigeration system of the refrigeration room to ensure the robustness of the control strategy under various conditions. The construction of the adaptive reinforcement learning model is as follows: Introduce a multi-scale state representation method to define the state representation, and transform the original state representation Decompose into feature representations at multiple scales , which is expressed as follows: ; in, represents the feature input at the mth scale, is the feature weight matrix of the current scale; Feature representation at multiple scales A hybrid strategy optimization mechanism is designed to handle high-dimensional continuous control problems. The hybrid strategy optimization mechanism is expressed as follows: ; in, represents the optimal strategy, represents the discount factor, represents the immediate reward, π represents the old strategy, T represents the total time duration, and t represents the current moment; A temperature adjustment mechanism is introduced to automatically adjust the exploration rate by monitoring system performance changes and learning rate , which is expressed as follows: ; in, and denote the minimum and maximum exploration rates, respectively. Indicates adjusting speed parameters.

2. The energy-saving control method for a centralized air-conditioning refrigeration system based on deep learning according to claim 1 is characterized in that: Various types of sensors are arranged in different spatial locations and devices, forming a multi-dimensional data collection matrix , which is expressed as follows: ; in, represents the temperature at position i, represents the humidity of the i-th device, represents the power consumption of the i-th device, Indicates the status of the i-th compressor, 0 means off, 1 means on, represents the fan speed of the i-th device, and n is the total number of sensors; The time synchronization process is represented as follows: ; in, represents the data of the i-th sensor processed in time synchronization, represents the raw data of the i-th sensor, is the time delay of the sensor, is the synchronization adjustment coefficient, It is a dynamic time offset item calculated based on historical data and is used to refine time synchronization; The preprocessing includes data calibration and noise processing, data cleaning and missing value filling and data standardization.

3. The energy-saving control method for a centralized air-conditioning refrigeration system based on deep learning according to claim 2 is characterized in that: Combined Temperature and Humidity Index Combining the characteristics of temperature and humidity, it is expressed as follows: ; The stability of the equipment state Reflects the stability of the equipment's operating status, as shown below: ; Where T represents the total number of windows and k represents the kth data point.

4. The energy-saving control method for a centralized air-conditioning refrigeration system based on deep learning according to claim 1 is characterized in that: The loss function of the deep learning model based on multi-task learning is expressed as follows: ; in, represents the loss function of the control task, represents the loss function of energy consumption prediction and anomaly detection tasks, represents the regularization term, which is used to prevent overfitting. , and represents the weight coefficient between tasks, Represents the model parameters of a deep learning model based on multi-task learning; Among them, the loss function of the control task is , which is expressed as follows: ; in, represents the immediate reward obtained at time step t, represents the discount factor at time step t, which is used to balance short-term and long-term benefits; The loss function for the energy consumption prediction and anomaly detection tasks is , which is expressed as follows: ; in, represents the energy consumption value predicted by the model, Indicates the actual energy consumption value, represents the regularization coefficient, Represents the second-order derivative of the predicted output with respect to the input feature, M represents the total number of features, and N represents the total number of samples; The hybrid regularization strategy is expressed as follows: ; in, represents the hybrid regularization term, express Regularization term, represents the kth feature weight Regularization term, K represents the number of features, and represents the corresponding regularization coefficient.

5. The energy-saving control method for a centralized air-conditioning refrigeration system based on deep learning according to claim 1 is characterized in that: The S4 specifically includes: S401, design the training objectives of the local models in each computer room, where the goal of the local model in each computer room is to minimize the local loss function , which is expressed as follows: ; in, and represents the data and label of the i-th training sample, represents the model parameters of computer room k, represents the regularization coefficient, Represents the personalized feature matrix of the computer room, represents element-wise product, represents the loss function, Indicates the actual operation data of the computer room. Indicates the operating characteristics of the computer room; S402: After each computer room completes local model training, an adaptive weight sharing mechanism is used to perform collaborative optimization on a global scale, as shown below: ; in, represents the global model parameters, represents the adaptive weight related to the data quality and model performance of computer room k; S403: Update the global model parameters The feedback mechanism is used to transmit the information back to each computer room. At the same time, a momentum correction mechanism is designed to adjust the degree of integration of new and old parameters. As the initial parameters for the next round of local training; wherein the momentum correction mechanism is expressed as follows: ; in, Represents the momentum factor, which controls the weight ratio of new and old parameters.

6. The energy-saving control method for a centralized air-conditioning refrigeration system based on deep learning according to claim 5 is characterized in that: The adaptive weight , calculated as follows: ; in, represents the adjustment parameter that controls the sensitivity of the weight distribution, represents the global model parameters of computer room k, represents the global model parameters of computer room j, represents the local loss function of computer room k, Represents the local loss function of computer room j.

7. The energy-saving control method for a centralized air-conditioning refrigeration system based on deep learning according to claim 1 is characterized in that: The prediction formula of the multi-order autoregressive model based on adaptive time series is expressed as follows: ; in, represents the predicted energy consumption value at time step t, represents the actual energy consumption value of the previous i time steps, represents the autoregressive coefficient, represents the moving average coefficient, represents the residual of the first j time steps, represents the offset term; Represents a multi-dimensional feature correction term, where represents the multidimensional feature vector at time step t, Represents a multi-dimensional feature correction item The corresponding weight vector; A multi-task loss function based on a multi-order autoregressive model of adaptive time series is designed. The multi-task loss function includes energy consumption prediction loss and anomaly detection loss, which is expressed as follows: ; in, and Represent the weight coefficients of energy consumption prediction error and anomaly detection loss, respectively. represents the threshold used to detect anomalies, represents the standard deviation of the residuals, which is used to weight the confidence of outliers. represents the actual energy consumption value at time step t; in, To minimize the prediction error, By adjusting the confidence level, higher detection weights are given to abnormal situations.

8. The energy-saving control method for a centralized air-conditioning refrigeration system based on deep learning according to claim 7 is characterized in that: In the training process of the multi-order autoregressive model based on adaptive time series, if the prediction error of a certain time period is large, the data of the current time period will be given the highest weight in subsequent training, wherein the weight adjustment formula is expressed as follows: ; in, represents the sample weight at time step t, represents the sensitivity parameter for weight adjustment; The design regularization term improves the stability of the multi-order autoregressive model based on adaptive time series when processing high-dimensional features. , which is expressed as follows: ; in, represents the weight vector of the i-th feature, , and They represent the regularization hyperparameters, Represents a small constant used to prevent the denominator from being zero.

9. The energy-saving control method for a centralized air-conditioning refrigeration system based on deep learning according to claim 1, characterized in that: The comprehensive loss function of the multi-objective optimization model It is expressed as follows: ; in, represents the energy consumption at time step t, Indicates the cooling efficiency, represents the standard deviation of temperature fluctuations, represents the rate of change of temperature, represents the risk assessment regularization term, which is used to evaluate potential system risks. , , , Respectively represent the weight coefficient of each target; wherein the targets include energy consumption, cooling efficiency and system stability; The risk assessment regularization term It is expressed as follows: ; in, represents the kth control parameter of the system, represents the risk weight coefficient, represents the second-order derivative of energy consumption to the control parameter, which is used to capture the sensitivity of the system. Indicates the sensitivity of temperature fluctuations to control parameters; The dynamic weight adjustment multi-objective optimization model loss function is set to automatically adjust the weight coefficient of each objective in the loss function according to the real-time status of the system in actual operation. , which is expressed as follows: ; in, represents the initial base weight, represents the adjustment coefficient, Represents the disturbance term at time step t, which is used to introduce a certain amount of randomness to prevent falling into the local optimal solution.

10. A refrigeration system energy-saving control system for a refrigeration room based on deep learning, characterized in that: The system comprises: The parameter collection module is used to deploy various types of sensors in the refrigeration system of the refrigeration room to monitor key parameters, perform time synchronization processing and preprocessing on the key parameters, and introduce the temperature and humidity combined index and equipment state stability to extract features from the preprocessed data; wherein the key parameters include temperature, humidity, energy consumption and equipment operating status; The energy consumption prediction and anomaly detection module is used to build a deep learning model based on multi-task learning, which divides the deep learning model training task into the task of real-time control and the task of energy consumption prediction and anomaly detection. At the same time, a hybrid regularization strategy is added to the deep learning model to prevent overfitting of the deep learning model training; the hybrid regularization strategy is combined with Regularization and sparse regularization; An operating parameter optimization module is used to design an adaptive reinforcement learning model based on multi-scale state representation and hybrid strategy optimization to optimize the operating parameters of the refrigeration system of the refrigeration room by training with the output of a deep learning model based on multi-task learning; A federated model building module is used to design a federated learning scheme based on an adaptive weight sharing mechanism to achieve collaborative optimization of the operating parameters of the refrigeration system of multiple distributed refrigeration rooms; The sudden anomaly optimization module is used to perform energy consumption prediction and anomaly detection on the refrigeration system of the refrigeration room according to the loss function of the energy consumption prediction and anomaly detection tasks when the refrigeration system of the refrigeration room has complex nonlinear behaviors and sudden anomalies. The module also uses an adaptive time series multi-order autoregressive model to perform energy consumption prediction and anomaly detection on the refrigeration system of the refrigeration room. In the training process, an adaptive weight adjustment mechanism is used. If the prediction error of a certain time period is large, the data of the current time period will be given a higher weight in subsequent training. At the same time, a regularization term is designed to improve the stability of the multi-order autoregressive model based on adaptive time series when processing high-dimensional features and reduce the risk of overfitting. The system optimization module is used to define a multi-objective optimization model to integrate multiple objectives and introduce a risk assessment regularization term to evaluate potential system risks. At the same time, dynamic weights are set to adjust the loss function of the multi-objective optimization model to ensure that the refrigeration system of the refrigeration room always focuses on the most important goal under different working conditions. Finally, feedback control based on reinforcement learning is set and the multi-objective optimization model is optimized. A disturbance assessment mechanism based on Monte Carlo simulation is introduced to simulate external disturbances. Then, the control strategy of the refrigeration system of the refrigeration room is evaluated and screened to ensure the robustness of the control strategy under various conditions. The construction of the adaptive reinforcement learning model is as follows: Introduce a multi-scale state representation method to define the state representation, and transform the original state representation Decompose into feature representations at multiple scales , which is expressed as follows: ; in, represents the feature input at the mth scale, is the feature weight matrix of the current scale; Feature representation at multiple scales A hybrid strategy optimization mechanism is designed to handle high-dimensional continuous control problems. The hybrid strategy optimization mechanism is expressed as follows: ; in, represents the optimal strategy, represents the discount factor, represents the immediate reward, π represents the old strategy, T represents the total time duration, and t represents the current moment; A temperature adjustment mechanism is introduced to automatically adjust the exploration rate by monitoring system performance changes and learning rate , which is expressed as follows: ; in, and denote the minimum and maximum exploration rates, respectively. Indicates adjusting speed parameters.