Cold and hot channel air volume optimization energy-saving method based on AI prediction

Abstract

The invention discloses a cold and hot channel air volume optimization energy-saving method based on AI prediction. The method comprises the following steps that S1, multi-dimensional data are collected and preprocessed; s2, inputting a double-branch convolution encoder, and extracting spatial migration thermal features; s3, inputting the spatial migration thermal characteristics into the spatial migration adversarial network model, and outputting a thermal load prediction map; s4, constructing an air volume adjustment multi-objective optimization model based on the thermal load prediction map, and searching an air volume configuration optimal solution by adopting a Harris eagle optimization algorithm; s5, feeding back an air volume configuration result to the space migration adversarial network model, and adjusting a training strategy; s6, the air volume configuration result is converted into an equipment control instruction, and actual operation feedback data is obtained; and S7, comparing the feedback data with the thermal load prediction map, if the difference exceeds a threshold value, returning to the step S3, and carrying out online updating on the spatial migration adversarial network model. According to the method, space migration modeling and an intelligent optimization algorithm are fused, and accurate prediction and dynamic control of the cold and hot channel air volume are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of energy-saving control, and particularly to an energy-saving method for optimizing the air volume of cold and hot channels based on AI prediction. Background Art

[0002] In modern data centers, in order to ensure the stable operation of high-performance computing devices in a high-density deployment environment, the cold and hot channel air volume regulation technology is widely used. This technology realizes the physical isolation and guiding control of cold and hot airflows by uniformly orienting the air inlet surfaces of devices towards the cold channel and the air outlet surfaces towards the hot channel, thereby improving the refrigeration efficiency and energy-saving effect. However, with the increasing volatility of server power consumption, the frequent changes in load scheduling, and the intelligent development of air-conditioning systems, the traditional static air volume configuration method has become difficult to meet the requirements of complex and changeable thermal environments. There is an urgent need for a dynamic air volume optimization technology with predictive, adaptive, and intelligent adjustment capabilities.

[0003] Most of the existing cold and hot channel air volume adjustment methods use rule-driven or simple feedback mechanisms to control the opening of air valves and the speed of fans, mainly relying on the temperature monitoring results after setting thresholds to trigger air volume adjustment. These methods usually ignore the coupled distribution relationship of heat in the spatial structure and lack the ability to predict future heat load trends. They often can only passively adjust when local temperature rise has occurred, resulting in frequent response lags, local overcooling, or overheating phenomena. In addition, traditional optimization algorithms such as genetic algorithms and particle swarm optimization are prone to falling into local optima when dealing with high-dimensional air volume configuration problems and lack the ability to coordinate multiple objectives in complex refrigeration scenarios, making it difficult to achieve an effective balance among temperature stability, minimum energy consumption, and control smoothness.

[0004] In recent years, with the rapid development of artificial intelligence technology, some studies have attempted to introduce deep learning models to model and predict the temperature control state of data centers. For example, structures such as recurrent neural networks (RNNs) and long short-term memory networks (LSTMs) are used to learn temperature time-series data to achieve a certain degree of heat load prediction ability. However, such models generally focus on the time series dimension and lack the ability to model the spatial topology structure of data centers and the characteristics of heat spatial migration, making it difficult to capture the conduction and coupling laws of heat distribution between physically adjacent racks, thus affecting the prediction accuracy and generalization ability. In addition, most of these models are mainly single-path input and cannot maintain prediction stability when the structure changes (such as device rearrangement and air duct reconstruction).

[0005] In terms of air volume optimization, traditional methods generally adopt single-objective optimization strategies and it is difficult to consider multiple operating constraints and energy efficiency objectives simultaneously. Between heat load prediction and air volume regulation, most methods still maintain a loose coupling structure, that is, the prediction only serves as an independent pre-module and cannot achieve the reverse effect of the optimization control result on the prediction model. The system lacks the ability of self-learning and evolution, and the model is prone to aging over time, resulting in a gradual weakening of the energy-saving effect.

[0006] Therefore, how to provide an energy-saving method for optimizing the air volume of cold and hot channels based on AI prediction is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0007] An object of the present invention is to propose an energy-saving method for optimizing the air volume of cold and hot channels based on AI prediction. The present invention integrates spatial migration modeling, adversarial prediction mechanism and Harris hawk optimization algorithm, systematically realizes the accurate prediction of heat load and multi-objective dynamic regulation of the air volume of cold and hot channels, constructs an intelligent closed-loop system of prediction-optimization-feedback-self-learning, and has the advantages of high energy-saving efficiency, strong model adaptability, and fast and accurate regulation response. It breaks through the technical bottlenecks of traditional methods in spatial coupling expression, regulation optimization and model evolution, and is applicable to the thermal management and energy consumption control systems of the new generation of green intelligent data centers.

[0008] An energy-saving method for optimizing the air volume of cold and hot channels based on AI prediction according to an embodiment of the present invention includes the following steps:

[0009] S1. Collect multi-dimensional data in the cold and hot channel areas and perform preprocessing;

[0010] S2. Input the preprocessed multi-dimensional data into a double-branch convolutional encoder, the double-branch convolutional encoder includes a main branch and an auxiliary branch, and fuse the high-dimensional features of the main branch and the auxiliary branch to extract spatial migration heat features;

[0011] S3. Input the spatial migration heat features into a spatial migration adversarial network model, optimize the heat load prediction result through the adversarial mechanism, and output a heat load prediction map;

[0012] S4. Based on the heat load prediction map, construct a multi-objective optimization model for air volume regulation, use the Harris hawk optimization algorithm to search for the optimal solution of air volume configuration, and output the air volume configuration result;

[0013] S5. Feed back the air volume configuration result to the spatial migration adversarial network model, adjust the training strategy of the spatial migration adversarial network model, and enhance the perception ability of the influence of control behavior;

[0014] S6. Convert the air volume configuration result into a device control instruction, adjust the opening of the corresponding channel air valve or the rotation speed of the fan, collect the temperature and energy consumption data after the air volume adjustment, and obtain the actual operation feedback data;

[0015] S7. Compare the actual operation feedback data with the heat load prediction graph. If the difference exceeds the threshold, return to step S3 and perform an online update on the spatial migration adversarial network model.

[0016] Optionally, the multi-dimensional data includes temperature, humidity, wind speed, rack power consumption, and damper opening.

[0017] Optionally, step S2 specifically includes:

[0018] S21. Rearrange the preprocessed multi-dimensional data in terms of time dimension and spatial dimension to construct a four-dimensional tensor X ∈ R T×H×W×D , where R represents the set of real numbers, T represents the length of the time series, H represents the height of the cold and hot channel regions in the spatial structure, W represents the width of the cold and hot channel regions in the spatial structure, and D represents the total number of channels;

[0019] S22. Input the four-dimensional tensor X into the main branch convolution path to extract the spatial heat features at the current moment and output the high-dimensional feature F main ;

[0020] S23. Retrieve the segment from the historical data that is most similar to the heat distribution pattern of the four-dimensional tensor X and input it into the auxiliary branch convolution path to output the high-dimensional feature F aux ;

[0021] S24. Construct an attention matrix and calculate the similarity of the features at each spatial position between the main branch and the auxiliary branch for position alignment preprocessing;

[0022] S25. Connect the high-dimensional feature F main of the main branch and the high-dimensional feature F aux of the auxiliary branch in the channel dimension and send them into a nested non-linear mapping structure to obtain an intermediate fusion representation:

[0023]

[0024] Among them, represents the intermediate fusion representation, GELU represents the activation function, W1 and W2 represent the weight matrices, b1 and b2 represent the bias terms, GELU represents the Gaussian error linear unit activation function, ReLU represents the non-linear activation function, [F main ∥ F aux represents the feature concatenation operation;

[0025] S26. Generate the final spatial migration heat features according to the cross-channel co-variation relationship:

[0026]

[0027] Among them, Represents the final fusion result of the d-th channel, Represents the intermediate fusion representation of the d-th channel, Represents the high-dimensional feature of the main branch of the d-th channel, Represents the high-dimensional feature of the auxiliary branch of the k-th channel, Represents the high-dimensional feature of the main branch of the k-th channel, Represents the high-dimensional feature of the auxiliary branch of the d-th channel, ∈ represents the smoothing factor, and ∥·∥2 represents the Euclidean norm.

[0028] Optionally, in S23, by constructing a retrieval mechanism based on the thermal distribution difference degree, the fragment most similar to the thermal distribution pattern of the four-dimensional tensor X is retrieved from the historical data, specifically including calculating the thermal distribution difference degree Y between the four-dimensional tensor X and the historical fragment i The thermal distribution difference degree between:

[0029]

[0030] where Z i Represents the thermal distribution difference degree between the four-dimensional tensor X and the i-th historical fragment Y i The height of the hot and cold channel regions in the spatial structure, and W represents the width of the hot and cold channel regions in the spatial structure, Represents the temperature gradient value of the four-dimensional tensor X at the spatial position (h, w), Represents the i-th historical fragment Y i The temperature gradient value at the spatial position (h, w), Represents the wind speed value of the four-dimensional tensor X at the spatial position (h, w), Represents the i-th historical fragment Y i The wind speed value at the spatial position (h, w), and α represents the weighting coefficient between the temperature gradient difference term and the wind speed difference term;

[0031] By calculating the thermal distribution difference degree for all historical fragments, the historical fragment with the minimum thermal distribution difference degree is selected as the most similar fragment and input into the auxiliary branch convolution path.

[0032] Optionally, the spatial migration adversarial network model includes a generator and a discriminator. The generator generates a heat load prediction map for each region at a future moment based on the input spatial migration thermal features. The discriminator then discriminates between the heat load prediction map output by the generator and the historical true heat load map, and continuously adjusts and optimizes the generator parameters through adversarial training to achieve dynamic modeling of cross-region heat transfer, and finally outputs a heat load prediction map reflecting the future heat load distribution.

[0033] Optionally, S4 specifically includes:

[0034] S41. Convert the heat load prediction map into a heat load matrix for each channel spatial unit, and extract the peak temperature, average temperature, temperature gradient change rate, and historical prediction error of each channel spatial unit;

[0035] S42. Set the control variables for air volume adjustment to construct a multi-objective optimization model for air volume adjustment, and establish three optimization objectives:

[0036] First, the minimum fluctuation of the outlet air temperature of each channel spatial unit;

[0037] Second, the lowest total fan energy consumption;

[0038] Third, the smoothest change in air volume;

[0039] S43. Construct a comprehensive fitness function to characterize the global performance of each candidate air volume configuration plan, and integrate three objectives: temperature error, energy consumption, and air volume smoothness:

[0040]

[0041] Among them, f i represents the comprehensive fitness value of the i-th candidate air volume configuration plan, ω1, ω2, and ω3 represent the weighting coefficients, D represents the total number of channels, represents the predicted outlet air temperature of the j-th channel spatial unit, represents the target temperature of the j-th channel spatial unit, M represents the number of fans, P k represents the energy consumption of the k-th fan, T represents the length of the time series, Q t represents the air volume distribution vector at the t-th moment, Q t+1 represents the air volume distribution vector at the (t + 1)-th moment, and ∥·∥2 represents the Euclidean norm;

[0042] S44. Based on the comprehensive fitness function, use the Harris hawk optimization algorithm to construct an air volume configuration search mechanism. Each hawk in the hawk group represents a candidate air volume configuration plan, and each candidate air volume configuration plan corresponds to a set of regional damper opening degrees and fan speed parameters;

[0043] S45. During the population iteration process, the Harris hawk optimization algorithm dynamically switches among the four-stage strategies of exploration, encirclement, dive, and strike, and adaptively balances between the global and local solutions to improve the search speed and stability of the optimal solution;

[0044] S46. After the iteration is completed, select the candidate air volume configuration plan with the smallest comprehensive fitness function value from the set of candidate air volume configuration plans as the optimal solution for air volume configuration, and output the air volume configuration result.

[0045] Optionally, the specific content of S5 includes:

[0046] S51. Associate the air volume configuration result with the heat load prediction map in terms of spatial position to generate a prediction-control comparison tensor set for marking the actual intervention direction of the control behavior on the thermal state of each region;

[0047] S52. Construct a control intervention label tensor, where the element values of the intervention label tensor are defined based on the coupling relationship between the predicted regional temperature change trend and the change direction of the corresponding control variable, and are input as additional training signals into the discriminator of the spatial transfer adversarial network model;

[0048] S53. Construct a reinforcement prediction control perception loss function \(L\) of the spatial transfer adversarial network model feedback :

[0049]

[0050] where \(L\) adv represents the original generator-discriminator adversarial loss, \(\eta\) represents the control perception weight coefficient, \(N\) represents the number of regions, \(\hat{T}_{p}\) represents the temperature of the \(p\)-th region predicted by the generator, \(u_{p}\) represents the air volume control value of the \(p\)-th region, \(\Delta T_{p}\) represents the temperature change value of the \(p\)-th region after the actual air volume adjustment is executed;

[0051] S54. Apply the reinforcement prediction control perception loss function \(L\) feedback to the update path of the generator of the spatial transfer adversarial network model, perform control perception-guided gradient descent optimization, and enhance the structural response memory of the spatial transfer adversarial network model for the air volume adjustment result;

[0052] S55. During the generator update process, construct a dynamic perception graph structure for re-weighting the perception connection strength between regions in each round of training:

[0053]

[0054] where \(G\) t \((p,q)\) represents the perception connection strength between the \(p\)-th region and the \(q\)-th region in the \(t\)-th round of training, \(\exp\) represents the exponential function, \(\gamma\) represents the temperature control sensitivity decay factor, \(\Delta T_{q}\) represents the temperature change value of the \(q\)-th region after the actual air volume adjustment is executed, \(u_{q}\) represents the air volume control value of the \(q\)-th region;

[0055] S56. Based on the updated generator structure and dynamic perception graph structure, train the new round of spatial transfer thermal features to enhance the perception ability of the spatial transfer adversarial network model for the impact of control behavior.

[0056] Optionally, the specific content of S6 includes:

[0057] S61. Receive the air volume configuration result, and based on the physical mapping relationship between the hot and cold channel areas and the control devices, establish a parameter binding table for the air volume configuration to the target control devices;

[0058] S62. According to the execution ability boundary and interface characteristics of the control object, discretize the continuous adjustment amount in the air volume configuration result, generate the percentage of damper opening and the fan speed level command parameters, and form a control command set;

[0059] S63. Send the control command set to the corresponding hot and cold channel control execution devices, so that the dampers or fans of each channel perform dynamic adjustment according to the air volume configuration result;

[0060] S64. During the execution of the control command, collect the temperature and energy consumption data after the air volume adjustment in real time to obtain the actual operation feedback data.

[0061] Optionally, the online update of the spatial migration adversarial network model specifically includes: constructing incremental training samples by introducing the actual operation feedback data, performing differential fusion processing on the actual operation feedback data and the original training data, retaining the structure and parameters of the original spatial migration adversarial network model, and performing local fine-tuning on the weights of the convolutional layers in the generator network and the discrimination strategy in the discriminator; during the online update process, freeze the underlying feature extraction layer and only train the high-level representation and output layer to ensure that the spatial migration adversarial network model adapts to the current environmental heat load distribution change without destroying the existing generalization ability; after the update is completed, automatically replace the old version of the spatial migration adversarial network model and restore the inference process to complete the closed-loop process from error detection to model evolution.

[0062] The beneficial effects of the present invention are as follows:

[0063] First, by constructing a multi-source sensing data-based hot and cold channel multi-dimensional information structure including temperature, humidity, wind speed, power consumption, damper opening, etc., and combining edge computing to complete data preprocessing, and further establishing a spatial topology structure according to the rack layout and thermal coupling relationship, the present invention realizes the accurate digital modeling of the hot and cold channel thermal environment, provides refined and structured input support for subsequent heat load analysis, and solves the problem of insufficient spatial thermal coupling expression ability of existing methods.

[0064] Second, the present invention uses a dual-branch convolutional encoder for spatial migration thermal feature extraction. The main branch captures the current hot and cold states, and the auxiliary branch automatically retrieves similar segments of the thermal distribution pattern in history, and fuses the features of both to construct a spatial migration thermal feature with cross-structure adaptation ability, effectively improving the modeling robustness in dynamic scenarios such as equipment rearrangement and load change. Compared with the traditional single-path structure, this feature extraction method has stronger generalization ability and expression ability in dealing with complex thermal distribution scenarios.

[0065] In addition, the present invention proposes a spatial migration adversarial network model as the core of thermal load prediction. The generator predicts the future thermal field distribution, and the discriminator makes difference judgments based on the historical real thermal map. The prediction strategy is continuously optimized through the adversarial training mechanism, which significantly enhances the model's ability to identify hot spot mutation areas in advance. The Harris Eagle optimization algorithm introduced on this basis has a four-stage dynamic strategy control mechanism, combined with a multi-objective model for air volume regulation (including temperature stability, minimum energy consumption, fluctuation control, etc.) to achieve the optimal solution search for high-dimensional air volume configuration space, overcoming the problems of local convergence and target imbalance of traditional optimization algorithms, and ensuring the dual goal of energy saving and temperature control.

[0066] Furthermore, the present invention returns the optimization results as feedback information to the heat load prediction model, and constructs a control-aware loss function to guide the model update, thus achieving deep coupling between prediction and control, and forming a prediction-control-feedback-retraining co-evolution mechanism. This mechanism not only improves the model's responsiveness to control behavior, but also has the ability to update online and dynamically adapt, solving the problem that traditional prediction models cannot adjust with environmental changes, and enabling the entire system to have long-term evolution and continuous optimization capabilities.

[0067] Finally, the present invention builds a complete control execution process, from discrete mapping of air volume configuration parameters, generation of control command sets, execution of equipment adjustment to real-time feedback collection, opening up the entire chain from high-dimensional intelligent optimization to actual control implementation, ensuring the engineering feasibility and system controllability of the optimization results. Overall, the present invention realizes the transformation of the hot and cold channel air volume optimization control system from passive response to active prediction, from static rules to intelligent scheduling, and from one-way control to closed-loop learning, significantly improving the energy efficiency and intelligence level of the data center temperature control system, and has good promotion value and application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0069] Figure 1 This is an overall flow chart of a cold and hot channel air volume optimization and energy saving method based on AI prediction proposed by the present invention;

[0070] Figure 2 A schematic diagram of the structure of a dual-branch convolutional encoder for extracting spatial migration thermal features in a cold and hot channel air volume optimization and energy-saving method based on AI prediction proposed by the present invention;

[0071] Figure 3It is a block diagram of the generator and discriminator of a spatial transfer adversarial network model for an AI - prediction - based hot - cold aisle air volume optimization and energy - saving method proposed by the present invention. Detailed implementation manners

[0072] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only schematically showing the basic structure of the present invention, so they only show the components related to the present invention.

[0073] Refer to Figures 1-3 , an AI - prediction - based hot - cold aisle air volume optimization and energy - saving method, which includes the following steps:

[0074] S1. Collect multi - dimensional data in the hot - cold aisle area and perform pre - processing;

[0075] S2. Input the pre - processed multi - dimensional data into a dual - branch convolutional encoder. The dual - branch convolutional encoder includes a main branch and an auxiliary branch, and fuse the high - dimensional features of the main branch and the auxiliary branch to extract spatial transfer thermal features;

[0076] S3. Input the spatial transfer thermal features into a spatial transfer adversarial network model, optimize the thermal load prediction result through an adversarial mechanism, and output a thermal load prediction map;

[0077] S4. Build a multi - objective optimization model for air volume regulation based on the thermal load prediction map, use the Harris hawk optimization algorithm to search for the optimal solution of air volume configuration, and output the air volume configuration result;

[0078] S5. Feed the air volume configuration result back to the spatial transfer adversarial network model, adjust the training strategy of the spatial transfer adversarial network model, and enhance the perception ability of the influence of control behaviors;

[0079] S6. Convert the air volume configuration result into a device control instruction, adjust the opening of the corresponding channel air valve or the rotation speed of the fan, collect the temperature and energy consumption data after the air volume adjustment, and obtain the actual operation feedback data;

[0080] S7. Compare the actual operation feedback data with the thermal load prediction map. If the difference exceeds the threshold, return to step S3 and perform online update on the spatial transfer adversarial network model.

[0081] The overall method proposed by the present invention takes AI prediction as the core, integrates spatial thermal feature modeling, thermal load prediction, air volume optimization regulation, and model feedback update to form a complete closed-loop system. Compared with the prior art, this method can not only accurately predict future thermal loads, but also dynamically configure the air volume according to the prediction results, realizing the intelligent matching of air volume and temperature control objectives, and avoiding the problems of regulation lag and energy consumption waste caused by relying on static rules or simple feedback control in traditional methods. At the same time, a feedback mechanism after air volume regulation is introduced, and through the online update of the prediction model, the system is given the ability of self-evolution, enabling it to still have strong robustness and adaptability in scenarios such as changes in equipment layout or dynamic fluctuations in load. The overall solution breaks the separation between prediction and control, realizes an intelligent air volume control strategy driven by data, and significantly improves the energy efficiency management level of the data center.

[0082] In this embodiment, the multi-dimensional data includes temperature, humidity, wind speed, rack power consumption, and damper opening.

[0083] By carefully collecting multi-dimensional data in the cold and hot aisle areas, including key parameters such as temperature, humidity, wind speed, power consumption, and damper opening, the present invention can comprehensively perceive the thermal environment and equipment operation status of the data center. The input constructed by this multi-source heterogeneous data fusion not only improves the data expression integrity for subsequent model training, but also solves the problem in the prior art that a single indicator cannot effectively reflect the state of the thermal field. The introduction of multi-dimensional data provides a more refined basic support for the extraction of spatial thermal features, enhances the response ability of the prediction model to environmental changes, lays a foundation for achieving high-accuracy thermal load prediction and high-precision air volume regulation, and significantly improves the reliability of system modeling and the actual control effect.

[0084] In this embodiment, S2 specifically includes:

[0085] S21. Rearrange the structure of the preprocessed multi-dimensional data according to the time dimension and the spatial dimension to construct a four-dimensional tensor X ∈ R T×H×W×D , where R represents the set of real numbers, T represents the length of the time series, H represents the height of the cold and hot aisle areas in the spatial structure, W represents the width of the cold and hot aisle areas in the spatial structure, and D represents the total number of channels;

[0086] S22. Input the four-dimensional tensor X into the main branch convolution path to extract the spatial thermal features at the current moment and output the high-dimensional feature F main ;

[0087] S23. Retrieve the segment from the historical data that is most similar to the thermal distribution pattern of the four-dimensional tensor X and input it into the auxiliary branch convolution path to output the high-dimensional feature F aux ;

[0088] S24. Construct an attention matrix and calculate the feature similarity of each spatial position between the main branch and the auxiliary branch for position alignment preprocessing;

[0089] S25. Feed the high-dimensional feature F of the main branch main and the high-dimensional feature F of the auxiliary branch aux into a nested non-linear mapping structure after concatenating them along the channel dimension to obtain an intermediate fusion representation:

[0090]

[0091] where, represents the intermediate fusion representation, GELU represents the activation function, W1 and W2 represent the weight matrices, b1 and b2 represent the bias terms, GELU represents the Gaussian error linear unit activation function, ReLU represents the non-linear activation function, [F main ∥F aux represents the feature concatenation operation;

[0092] S26. Generate the final spatial migration heat feature according to the cross-channel co-variation relationship:

[0093]

[0094] where, represents the final fusion result of the d-th channel, represents the intermediate fusion representation of the d-th channel, represents the high-dimensional feature of the main branch of the d-th channel, represents the high-dimensional feature of the k-th channel of the auxiliary branch, represents the high-dimensional feature of the main branch of the k-th channel, represents the high-dimensional feature of the d-th channel of the auxiliary branch, ∈ represents the smoothing factor, ∥·∥2 represents the Euclidean norm.

[0095] By introducing a dual-branch convolutional encoder structure, the main branch is responsible for modeling the current state, the auxiliary branch is responsible for referencing the historical pattern, and a fusion mechanism is designed to extract the spatial migration heat feature. This method effectively improves the stability of feature representation and the cross-structure generalization ability. Especially in scenarios such as dynamic adjustment of hot and cold channel structures, equipment rearrangement, and local cooling failure, this structure can automatically guide the learning direction of the main branch according to the similar heat distribution in history to avoid prediction deviation problems. At the same time, through the non-linear nested structure, the deep fusion of the main and auxiliary features is realized, enhancing the expression ability for complex heat conduction paths, enabling the subsequent prediction network to obtain more spatially aware inputs, and effectively improving the robustness and accuracy of the system in a dynamic thermal environment.

[0096] In this embodiment, in S23, by constructing a retrieval mechanism based on the thermal distribution difference degree, the fragment most similar to the thermal distribution pattern of the four-dimensional tensor X is retrieved from the historical data, which specifically includes calculating the thermal distribution difference degree Y between the four-dimensional tensor X and the historical fragment i The thermal distribution difference degree between

[0097]

[0098] where Z i represents the thermal distribution difference degree between the four-dimensional tensor X and the i-th historical fragment Y i H represents the height of the hot and cold channel regions in the spatial structure, W represents the width of the hot and cold channel regions in the spatial structure, represents the temperature gradient value of the four-dimensional tensor X at the spatial position (h, w), represents the temperature gradient value of the i-th historical fragment Y i at the spatial position (h, w), represents the wind speed value of the four-dimensional tensor X at the spatial position (h, w), represents the wind speed value of the i-th historical fragment Y i at the spatial position (h, w), and α represents the weighting coefficient between the temperature gradient difference term and the wind speed difference term;

[0099] By calculating the thermal distribution difference degree for all historical fragments, the historical fragment with the minimum thermal distribution difference degree is selected as the most similar fragment and input into the auxiliary branch convolution path.

[0100] By constructing a historical fragment retrieval mechanism based on the thermal distribution difference degree, the present invention can effectively identify the historical reference data closest to the current thermal state, thereby providing the thermal distribution information matching the current operating state for the auxiliary branch. This difference degree simultaneously considers two key dimensions of temperature gradient and wind speed change, fuses the structural features and dynamic behavior features for matching, and improves the accuracy of retrieval. Compared with the traditional static sample selection method, this method can realize intelligent dynamic sample recommendation, introduce more valuable historical scenarios in the process of spatial feature migration for the auxiliary branch, enhance the sensitivity of the system to the change of thermal coupling relationship, and improve the overall feature extraction effect and the stability of the prediction model.

[0101] In this embodiment, the spatial transfer adversarial network model includes a generator and a discriminator. The generator generates a thermal load prediction map for each region at the future moment based on the input spatial transfer thermal features. The discriminator then discriminates between the thermal load prediction map output by the generator and the historical true thermal load map. By adversarial training, the parameters of the generator are continuously adjusted and optimized to realize the dynamic modeling of cross-region heat transfer, and finally a thermal load prediction map reflecting the future thermal load distribution is output.

[0102] By constructing a spatial transfer adversarial network model with a generator-discriminator adversarial structure, high-precision modeling of future heat loads and prediction of abnormal regions are achieved. The generator is responsible for predicting the future heat distribution in each region, and the discriminator verifies and discriminates based on historical real heat maps. Through the adversarial mechanism, the two continuously game and optimize, effectively avoiding the problem of prediction distortion of traditional models for hot spot mutations and abnormal points. At the same time, this structure has the ability to model region by region and learn cross-regional heat interference, supports dynamic expression of complex spatial heat fields, and effectively improves the prediction ability and control basis of the system under extreme conditions such as sudden heat load fluctuations and uneven heat generation of equipment.

[0103] In this embodiment, step S4 specifically includes:

[0104] S41. Convert the heat load prediction map into a heat load matrix of each channel spatial unit, and extract the peak temperature, average temperature, temperature gradient change rate, and historical prediction error of each channel spatial unit;

[0105] S42. Set the control variables for air volume adjustment to construct a multi-objective optimization model for air volume adjustment, and establish three optimization objectives:

[0106] The first is that the air outlet temperature fluctuation of each channel spatial unit is minimized;

[0107] The second is that the total fan energy consumption is the lowest;

[0108] The third is that the air volume change is the smoothest;

[0109] S43. Construct a comprehensive fitness function to characterize the global performance of each candidate air volume configuration plan, and integrate three objectives: temperature error, energy consumption, and air volume smoothness:

[0110]

[0111] Among them, f i represents the comprehensive fitness value of the i-th candidate air volume configuration plan, ω1, ω2, and ω3 represent weighting coefficients, D represents the total number of channels, represents the predicted air outlet temperature of the j-th channel spatial unit, represents the target temperature of the j-th channel spatial unit, M represents the number of fans, P k represents the energy consumption of the k-th fan, T represents the time series length, Q t represents the air volume distribution vector at the t-th moment, Q t+1 represents the air volume distribution vector at the (t + 1)-th moment, and ∥·∥2 represents the Euclidean norm;

[0112] S44. Based on the comprehensive fitness function, a Harris hawk optimization algorithm is used to construct an air volume configuration search mechanism. Each hawk in the hawk group represents a candidate air volume configuration plan, and each candidate air volume configuration plan corresponds to a set of regional air valve opening degrees and fan speed parameters;

[0113] S45. During the population iteration process, the Harris hawk optimization algorithm dynamically switches among the four-stage strategies of exploration, encirclement, swooping, and striking, adaptively balances between the global and local solutions, and improves the search speed and stability of the optimal solution;

[0114] S46. After the iteration is completed, the candidate air volume configuration plan with the smallest comprehensive fitness function value is selected from the candidate air volume configuration plan set as the optimal solution of the air volume configuration, and the air volume configuration result is output.

[0115] By constructing a multi-objective optimization model for air volume regulation and introducing the Harris hawk optimization algorithm for solution, the limitation of the traditional method of regulating air volume based on a single objective is broken through. The present invention can comprehensively balance among the regional temperature stability, the total system energy consumption, and the smoothness of air volume adjustment. The Harris hawk optimization algorithm has a four-stage search mechanism, can dynamically switch between the global and local, improves the global optimal search ability of air volume configuration, and avoids falling into local optimal solutions. Combining the comprehensive fitness function with the coupling relationship modeling in the high-dimensional space makes the air volume control more targeted and robust, and significantly improves the accuracy and execution efficiency of the energy-saving regulation strategy.

[0116] In this embodiment, the specific content of S5 is as follows:

[0117] S51. The air volume configuration result is spatially associated with the heat load prediction map to generate a prediction-control comparison tensor set, which is used to mark the actual intervention direction of the control behavior on the thermal state of each region;

[0118] S52. A control intervention label tensor is constructed. The element values of the intervention label tensor are defined based on the coupling relationship between the predicted regional temperature change trend and the change direction of the corresponding control variable, and are input as an additional training signal into the discriminator of the spatial transfer adversarial network model;

[0119] S53. Construct the enhanced prediction control perception loss function L of the spatial transfer adversarial network model feedback :

[0120]

[0121] where L adv represents the original generator-discriminator adversarial loss, η represents the control perception weight coefficient, N represents the number of regions, represents the temperature of the p-th region predicted by the generator, represents the air volume control value of the p-th region, Represents the temperature change value of the p-th area after the actual air volume adjustment is performed;

[0122] S54. Apply the enhanced prediction control perception loss function L feedback To the update path of the generator of the spatial migration adversarial network model, perform control perception-guided gradient descent optimization to enhance the structural response memory of the spatial migration adversarial network model for the air volume adjustment result;

[0123] S55. During the generator update process, construct a dynamic perception graph structure for re-weighting the perception connection strength between regions in each round of training:

[0124]

[0125] Where, G t (p, q) represents the perception connection strength between the p-th area and the q-th area in the t-th round of training, exp represents the exponential function, γ represents the temperature control sensitivity attenuation factor, Represents the temperature change value of the q-th area after the actual air volume adjustment is performed, Represents the air volume control value of the q-th area;

[0126] S56. Based on the updated generator structure and dynamic perception graph structure, train the new round of spatial migration thermal features to enhance the perception ability of the spatial migration adversarial network model for the impact of control behaviors.

[0127] By feeding back the air volume configuration result to the thermal load prediction model and constructing a control perception training mechanism, the deep linkage between prediction and control is realized. The present invention introduces control intervention labels and a control perception loss function, and uses the deviation information of the actual control behavior on the prediction result for discriminator training, effectively enhancing the model's response perception ability to control behaviors. By strengthening the dynamic coupling between the prediction path and the control variables, the model gradually establishes the structural memory between prediction - control - impact, effectively improving the feedforward adjustment ability of the prediction strategy for execution behaviors, thereby improving the prediction accuracy and control behavior consistency, and forming an adaptive evolutionary intelligent control path.

[0128] In this embodiment, the specific content of S6 includes:

[0129] S61. Receive the air volume configuration result, and based on the physical mapping relationship between the cold and hot channel areas and the control equipment, establish a parameter binding table for the air volume configuration to the control target equipment;

[0130] S62. According to the execution ability boundary and interface characteristics of the control object, perform discretization mapping on the continuous adjustment quantity in the air volume configuration result to generate air valve opening percentage and fan speed level command parameters, and form a control command set;

[0131] S63. Send the control command set to the corresponding cooling and heating channel control execution devices, so that each channel air valve or fan is dynamically adjusted according to the air volume configuration result;

[0132] S64. During the execution of the control command, collect the temperature and energy consumption data after the air volume adjustment in real time to obtain the actual operation feedback data.

[0133] By binding the optimized air volume configuration result with the control command of the execution device and performing discrete mapping in combination with the control boundary conditions and actual execution characteristics, seamless implementation from the AI optimization result to the actual device control behavior is achieved. In the process of parameter generation, the present invention fully considers the ability boundary of the execution device and the interface adaptation mechanism to ensure that the adjustment commands for the air valve opening and fan speed not only meet the control objectives but also have engineering feasibility. In addition, by collecting control feedback data in real time during the execution process, the system can form a high-frequency execution-feedback closed loop, which helps to continuously correct the deviation between the optimization target and the execution behavior and improve the control reliability and system intelligence level.

[0134] In this embodiment, the online update of the spatial transfer adversarial network model specifically includes: constructing incremental training samples by introducing the actual operation feedback data, performing differential fusion processing on the actual operation feedback data and the original training data, retaining the original spatial transfer adversarial network model structure and parameter basis, and performing local fine-tuning on the convolutional layer weights in the generator network and the discrimination strategy in the discriminator; during the online update process, the underlying feature extraction layer is frozen, and only the high-level representation and output layer are trained to ensure that the spatial transfer adversarial network model adapts to the current environmental heat load distribution change without destroying the existing generalization ability; after the update is completed, the old version of the spatial transfer adversarial network model is automatically replaced and the inference process is restored to complete the closed-loop process from error detection to model evolution.

[0135] The present invention realizes the self-evolution and continuous adaptation ability of the model by constructing an online update mechanism for the spatial transfer adversarial network model. During the operation process, the system constructs incremental training samples by collecting actual feedback data and performs local parameter fine-tuning on the key modules in the generator and discriminator structures, effectively avoiding the problem of decreased prediction accuracy caused by model aging or environmental changes. This update mechanism freezes the underlying structure and only trains the high-level representation layer, effectively taking into account both the update efficiency and the model stability. Finally, the continuous evolution and performance maintenance of the model under different loads, structures and environmental conditions are realized, significantly improving the long-term stable operation ability of the intelligent air volume control system.

[0136] Example 1:

[0137] To verify the feasibility of the present invention in implementation, the present invention is applied to the energy-saving control system of the cold and hot aisles in a certain ultra-large data center. This data center adopts a typical layout of cold aisle air intake and hot aisle air exhaust, deploys more than 1000 server racks, and sets up a air volume regulation system composed of multiple groups of independent air valves and variable-frequency fans. The original system adopts a static threshold control strategy, mainly based on temperature overrun for responsive regulation, and there are problems such as control delay, rough regulation, high energy consumption, uneven cold and heat distribution, etc. Especially during the period when the high-concurrency task load changes frequently, instantaneous overheating or energy consumption redundancy is likely to occur in the local rack area, affecting the system stability and energy efficiency.

[0138] In this embodiment, first, various types of sensors such as temperature, humidity, wind speed, rack power consumption, and air valve opening are deployed at key positions in the cold and hot aisles to collect multi-dimensional operation data in the cold and hot aisles in real time. Through the edge computing node, these original data are processed for standardization, denoising, and time series alignment to construct a spatial structure thermal field data set. Using the dual-branch convolutional encoder structure proposed by the present invention, the current thermal distribution state is matched and fused with data segments with similar thermal distribution patterns in history, and high-dimensional spatial thermal features that can reflect the structural migration characteristics are extracted. This feature is used as the input and fed into the spatial migration adversarial network model constructed by the present invention. The generator of the model predicts the thermal load change trend within the next ten minutes, and the discriminator discriminates and trains the predicted map of the generator and the historical real thermal map, so as to dynamically optimize the prediction structure.

[0139] The prediction result is then used as the input and fed into the air volume regulation multi-objective optimization model of the present invention. The model constructs a trade-off function among the three objectives of minimizing the fluctuation of the outlet air temperature, minimizing the fan energy consumption, and stabilizing the air volume fluctuation, and uses the Harris hawk optimization algorithm to search for the global optimal air volume configuration scheme. This optimization strategy effectively improves the convergence efficiency and the quality of the solution in the high-dimensional air volume space through the four-stage adaptive search strategy of the hawk, and the output result covers the air valve opening of each aisle and the fan speed level. Finally, the system converts this optimal configuration result into an executable instruction and automatically issues it to the device layer control unit to complete the air volume regulation.

[0140] During the operation of the system, the actual temperature feedback is continuously compared with the predicted thermal map. Once it is detected that the prediction deviation exceeds the threshold, the system automatically calls back the spatial migration adversarial network for local online update and fuses new data for periodic fine-tuning, so as to ensure that the model can adapt to complex thermal load changes in the long term and form a self-evolving closed-loop structure of prediction-control-feedback-re-prediction.

[0141] After 14 consecutive days of test data statistics, under high-load scenarios, compared with the traditional fixed air volume strategy, the system of the present invention reduces the overall fan energy consumption by about 21.4%, the average standard deviation of the outlet air temperature per channel decreases by 36.8%, the number of local overheating events is reduced by 78 times, the average delay of the system temperature control response is shortened to 6.2 seconds, and the prediction accuracy (based on the regional average temperature difference of less than ±0.5°C) is increased to 92.7%. In addition, through the model feedback update mechanism, the control behavior correction rate of the air volume control system after one week of operation is increased by about 18.5%, indicating that the system has good adaptive learning ability.

[0142] To further demonstrate the implementation effect, the following Table 1 lists the key performance indicators of the present invention and the traditional method in a typical test cycle:

[0143] Table 1 Performance comparison data of the solution of the present invention and the traditional solution

[0144]

[0145]

[0146] From the performance data comparison listed in Table 1, it can be seen that the intelligent air volume optimization system proposed in the present invention has significant advantages over the traditional control method in multiple key performance indicators. First, in terms of the total fan energy consumption of the system, the average daily power consumption of the present invention is 843.6kWh, while the traditional solution is as high as 1072.8kWh, with an energy saving of 21.4%. This shows that the prediction-driven air volume optimization control strategy can effectively reduce the ineffective air volume output and reduce the fan operating load.

[0147] In terms of outlet air temperature stability, the average fluctuation of the present invention is ±0.34°C, which is 36.8% lower than the ±0.54°C of the traditional solution. This shows that by accurately predicting future heat loads and dynamically adjusting the air volume configuration, the problem of equipment thermal shock caused by excessive temperature fluctuations can be effectively avoided. At the same time, local overheating events have dropped significantly from 99 times per day in the traditional system to 21 times, a reduction of about 78.8%, indicating that the system can respond to high-heat areas in a timely manner, provide targeted regulation, and ensure a more balanced overall thermal environment.

[0148] In terms of control response delay, the average response time of the intelligent optimization system is 6.2 seconds, which is much lower than the 17.4 seconds of the traditional system, shortened by more than 60%, greatly improving the real-time control capability of the system, which is conducive to coping with scenarios with rapid load changes. In terms of prediction accuracy, taking the difference between the predicted temperature and the actual value within the range of ±0.5℃ as the standard, the prediction accuracy of the present invention reaches 92.7%, which is 21.8% higher than the 76.1% of the traditional method, which fully verifies the advancement of the spatial migration adversarial network model in thermal load prediction.

[0149] In addition, during continuous operation, the intelligent system can automatically update the prediction model according to actual feedback. In the 7-day test, the feedback correction rate reached 18.5%, while the traditional system does not have such capabilities. This result further demonstrates the significant advantages of the present invention in model adaptive learning and long-term stable operation, reflecting the intelligent evolution characteristic of becoming more accurate with use.

[0150] Thus, from the six dimensions of energy consumption control, temperature stability, local heat dissipation capacity, response speed, prediction accuracy, and system adaptability, the air volume optimization and energy-saving method of the present invention exhibits excellent technical performance and obvious energy-saving effects in the actual data center application, and has strong engineering promotion value and industrial application prospects.

[0151] The above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, should be covered by the protection scope of the present invention.

Claims

1. An energy-saving method for optimizing the air volume in hot and cold channels based on AI prediction, characterized in that It includes the following steps: S1. Collect multi-dimensional data in the hot and cold channel areas and perform preprocessing; S2. Input the preprocessed multi-dimensional data into a dual-branch convolutional encoder, where the dual-branch convolutional encoder includes a main branch and an auxiliary branch, fuse the high-dimensional features of the main branch and the auxiliary branch, and extract spatially migrated thermal features; S3. Input the spatially migrated thermal features into a spatial migration adversarial network model, optimize the thermal load prediction result through an adversarial mechanism, and output a thermal load prediction map; S4. Based on the thermal load prediction map, construct a multi-objective optimization model for air volume regulation, use the Harris hawk optimization algorithm to search for the optimal solution of air volume configuration, and output the air volume configuration result; S5. Feed the air volume configuration result back to the spatial migration adversarial network model, adjust the training strategy of the spatial migration adversarial network model, and enhance the perception ability of the impact of control actions; S6. Convert the air volume configuration result into device control instructions, adjust the opening of the corresponding channel air valve or the rotation speed of the fan, collect the temperature and energy consumption data after the air volume adjustment, and obtain the actual operation feedback data; S7. Compare the actual operation feedback data with the thermal load prediction map. If the difference exceeds the threshold, return to step S3 and perform online update on the spatial migration adversarial network model.

2. The energy-saving method for optimizing the air volume of cold and hot channels based on AI prediction according to claim 1, characterized in that The multi-dimensional data includes temperature, humidity, wind speed, rack power consumption, and air valve opening.

3. The energy-saving method for optimizing the air volume of hot and cold channels based on AI prediction according to claim 1, wherein Specifically, S2 includes: S21. Rearrange the preprocessed multi-dimensional data in terms of the time dimension and the spatial dimension to construct a four-dimensional tensor \(X\in\mathbb{R}\) T×H×W×D , where \(\mathbb{R}\) represents the set of real numbers, \(T\) represents the length of the time series, \(H\) represents the height of the hot and cold channel regions in the spatial structure, \(W\) represents the width of the hot and cold channel regions in the spatial structure, and \(D\) represents the total number of channels; S22. Input the four-dimensional tensor X into the main branch convolution path to extract the spatial thermal features at the current moment and output the high-dimensional feature F of the main branch main ; S23. Retrieve the segment from historical data that is most similar to the thermal distribution pattern of the four-dimensional tensor X, and input it into the auxiliary branch convolution path to output the high-dimensional feature F of the auxiliary branch aux ; S24. Construct an attention matrix and calculate the feature similarity of each spatial position of the main branch and the auxiliary branch for position alignment preprocessing; S25. Feed the high-dimensional feature F of the main branch main and the high-dimensional feature F of the auxiliary branch aux into the nested non-linear mapping structure after concatenating them in the channel dimension to obtain an intermediate fusion representation: Among them, represents the intermediate fusion representation, GELU represents the activation function, W1 and W2 represent the weight matrices, b1 and b2 represent the bias terms, GELU represents the Gaussian error linear unit activation function, ReLU represents the non-linear activation function, [F main ∥F aux represents the feature concatenation operation; S26. Generate the final spatially migrated thermal feature according to the cross-channel co-variation relationship: Among them, represents the final fusion result of the d-th channel, represents the intermediate fusion representation of the d-th channel, represents the high-dimensional feature of the main branch of the d-th channel, represents the high-dimensional feature of the auxiliary branch of the k-th channel, represents the high-dimensional feature of the main branch of the k-th channel, represents the high-dimensional feature of the auxiliary branch of the d-th channel, ∈ represents the smoothing factor, and ∥·∥2 represents the Euclidean norm.

4. A method for optimizing energy conservation of air volume in hot and cold channels based on AI prediction according to claim 3, characterized in that, In S23, by constructing a retrieval mechanism based on the thermal distribution difference degree, the fragment most similar to the thermal distribution pattern of the four-dimensional tensor X is retrieved from the historical data, specifically including calculating the thermal distribution difference degree Y between the four-dimensional tensor X and the historical fragment i The thermal distribution difference degree between Among them, Z i represents the thermal distribution difference degree between the four-dimensional tensor X and the i-th historical segment Y i , H represents the height of the hot and cold channel region in the spatial structure, W represents the width of the hot and cold channel region in the spatial structure, represents the temperature gradient value of the four-dimensional tensor X at the spatial position (h, w), represents the i-th historical segment Y i at the temperature gradient value at the spatial position (h, w), represents the wind speed value of the four-dimensional tensor X at the spatial position (h, w), represents the i-th historical segment Y i at the wind speed value at the spatial position (h, w), and α represents the weighting coefficient between the temperature gradient difference term and the wind speed difference term; By calculating the thermal distribution difference degree for all historical segments, select the historical segment with the smallest thermal distribution difference degree as the most similar segment and input it into the auxiliary branch convolutional path.

5. A method for optimizing energy conservation of air volume in hot and cold channels based on AI prediction according to claim 1, characterized in that The spatial migration adversarial network model includes a generator and a discriminator. The generator generates a thermal load prediction map of each region at a future moment based on the input spatially migrated thermal features. The discriminator discriminates between the thermal load prediction map output by the generator and the historical true thermal load map, and continuously adjusts and optimizes the generator parameters through adversarial training to realize the dynamic modeling of cross-region heat transfer, and finally outputs a thermal load prediction map reflecting the future thermal load distribution.

6. The energy-saving method for optimizing the air volume of cold and hot channels based on AI prediction according to claim 1, wherein, Specifically, S4 includes: S41. Convert the thermal load prediction map into a thermal load matrix of each channel spatial unit, and extract the peak temperature, average temperature, temperature gradient change rate, and historical prediction error of each channel spatial unit; S42. Set the control variables for air volume regulation to construct a multi-objective optimization model for air volume regulation, and establish three optimization objectives: The first is the minimum fluctuation of the outlet air temperature of each channel spatial unit; The second is the lowest total fan energy consumption; The third is the smoothest air volume change; S43. Construct a comprehensive fitness function to characterize the global performance of each candidate air volume configuration plan, and fuse the three objectives of temperature error, energy consumption, and air volume smoothness; Among them, f i represents the comprehensive fitness value of the i-th candidate air volume configuration scheme, ω1, ω2, and ω3 represent the weighting coefficients, D represents the total number of channels, represents the predicted air outlet temperature of the j-th channel space unit, represents the target temperature of the j-th channel space unit, M represents the number of fans, P k represents the energy consumption of the k-th fan, T represents the length of the time series, Q t represents the air volume distribution vector at the t-th moment, Q t+1 represents the air volume distribution vector at the (t + 1)-th moment, ∥·∥2 represents the Euclidean norm; S44. Based on the comprehensive fitness function, use the Harris hawk optimization algorithm to construct an air volume configuration search mechanism. Each hawk in the hawk group represents a candidate air volume configuration plan, and each candidate air volume configuration plan corresponds to a set of regional air valve opening and fan speed parameters; S45. During the population iteration process, the Harris hawk optimization algorithm dynamically switches between the four-stage strategies of exploration, encirclement, dive, and strike, adaptively balances between the global and local solutions, and improves the search speed and stability of the optimal solution. S46. After the iteration is completed, the candidate air volume configuration plan with the smallest comprehensive fitness function value is selected from the candidate air volume configuration plan set as the optimal solution of the air volume configuration, and the air volume configuration result is output.

7. A method for optimizing energy conservation of air volume in hot and cold channels based on AI prediction according to claim 1, characterized in that The specific content of S5 is as follows: S51. Associate the air volume configuration result with the thermal load prediction map in terms of spatial position to generate a prediction-control comparison tensor set, which is used to mark the actual intervention direction of the control behavior on the thermal state of each region. S52. Construct a control intervention label tensor, the element values of which are defined based on the coupling relationship between the predicted regional temperature change trend and the change direction of the corresponding control variable, and are input as an additional training signal to the discriminator of the spatial transfer adversarial network model. S53. Construct the enhanced predictive control perception loss function L of the spatial migration adversarial network model feedback :[[]]END]] Among them, L adv represents the original generator-discriminator adversarial loss, η represents the control perception weight coefficient, N represents the number of regions, represents the temperature of the p-th region predicted by the generator, represents the air volume control value of the p-th region, represents the temperature change value of the p-th region after the actual air volume adjustment is executed; S54. Apply the enhanced predictive control perception loss function L feedback to the update path of the generator of the spatial transfer adversarial network model, perform control perception-guided gradient descent optimization, and enhance the structural response memory of the spatial transfer adversarial network model for the air volume adjustment result; S55. During the update process of the generator, construct a dynamic perception graph structure to re-weight the perception connection strength between regions in each round of training. Among them, G t (p, q) represents the perceived connection strength between the p-th area and the q-th area in the t-th round of training. exp represents the exponential function, and γ represents the temperature control sensitivity decay factor. represents the temperature change value of the q-th area after the actual air volume adjustment is performed. represents the air volume control value of the q-th area; S56. Based on the updated generator structure and dynamic perception graph structure, train the new round of spatial transfer thermal features to enhance the perception ability of the spatial transfer adversarial network model to the impact of control behavior.

8. A method for optimizing energy conservation of air volume in hot and cold channels based on AI prediction according to claim 1, characterized in that, The specific content of S6 is as follows: S61. Receive the air volume configuration result, and based on the physical mapping relationship between the cold and hot channel regions and the control devices, establish a parameter binding table for the air volume configuration to the control target devices. S62. According to the execution ability boundary and interface characteristics of the control object, discretize the continuous adjustment amount in the air volume configuration result to generate the percentage of damper opening and the command parameters of the fan speed level, and form a control command set. S63. Send the control command set to the corresponding cold and hot channel control execution devices, so that the dampers or fans of each channel are dynamically adjusted according to the air volume configuration result. S64. During the execution of the control command, collect the temperature and energy consumption data after the air volume adjustment in real time to obtain the actual operation feedback data.

9. A method for optimizing energy conservation of air volume in hot and cold channels based on AI prediction according to claim 1, characterized in that, The online update of the spatial transfer adversarial network model specifically includes: constructing incremental training samples by introducing the actual operation feedback data, performing differential fusion processing on the actual operation feedback data and the original training data, retaining the structure and parameters of the original spatial transfer adversarial network model, and locally fine-tuning the weights of the convolutional layers in the generator network and the discrimination strategy in the discriminator; during the online update process, freeze the underlying feature extraction layer and only train the high-level representation and output layer to ensure that the spatial transfer adversarial network model adapts to the current environmental thermal load distribution change without destroying the existing generalization ability; after the update is completed, automatically replace the old version of the spatial transfer adversarial network model and resume the inference process to complete the closed-loop process from error detection to model evolution.

Images