Multi-data center waste heat recycling optimization method and system based on federated learning

Abstract

The invention provides a multi-data center waste heat recycling optimization method and system based on federated learning, and relates to the field of data center waste heat recycling. According to the method, feature extraction and modeling are performed on multi-source data through a federated learning framework, a space-time neural network prediction model is trained on the premise of ensuring data privacy security of each data center, and future waste heat supply of each data center and building thermal load requirements of surrounding areas are obtained; and based on the waste heat recycling optimization model, performing system modeling and matching optimization on the waste heat supply of each data center and the heat load demand of the buildings in the surrounding area in the future so as to obtain an optimal waste heat distribution strategy. According to the method, the future waste heat supply of each data center and the building heat demand of the surrounding area are predicted through the federated learning framework, the waste heat recycling optimization in the encryption state is realized on the premise of ensuring that the data privacy is not leaked, and the problems of dependence on data centralized storage and privacy leakage risk in traditional centralized modeling are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data center waste heat recycling, and particularly relates to a multi-data center waste heat recycling optimization method and system based on federated learning. BACKGROUND

[0002] As a high energy consumption facility, data centers consume a large amount of electricity and generate a large amount of waste heat. If these waste heat can be effectively recycled, the energy efficiency can be significantly improved and carbon emissions can be reduced. The traditional data center waste heat utilization method mainly focuses on the internal of a single data center, and its waste heat utilization is limited by its own heat demand and equipment conditions. The current common method is to use waste heat to heat domestic water or office area heating in the data center, which is insufficient in depth and breadth of waste heat utilization. This single utilization mode not only cannot fully tap the potential value of waste heat, but also causes the waste heat production of a single data center to be unstable due to load fluctuation, which is difficult to match the dynamic demand of the regional heating network, resulting in energy waste or insufficient heating. Therefore, through the waste heat recycling optimization between multiple data centers, the waste heat sharing and heat load adjustment across data centers can be better realized, the waste heat utilization rate can be improved, and resource waste can be reduced.

[0003] The existing multi-data center joint waste heat recycling optimization method mainly includes a centralized method, which collects the operation data and waste heat information of each data center, and then uniformly distributes the waste heat of multiple data centers. For example, in some regions, the waste heat of multiple data centers is collected and aggregated, and then uniformly delivered to the regional heating network through a large heat exchange device to provide heating services for surrounding buildings.

[0004] However, this centralized method faces significant data privacy problems in actual application. Each data center is reluctant to directly share original operation data or load information due to business and security considerations, as it may involve key business indicators or operation strategies, which poses a risk of privacy leakage, which easily forms a data island (data island leads to low coordination efficiency) and a privacy barrier. That is, the existing multi-data center joint waste heat recycling optimization method cannot balance privacy protection and global waste heat recycling optimization. SUMMARY

[0005] (I) Technical problems solved

[0006] In view of the deficiencies of the prior art, the present application provides a multi-data center waste heat recycling optimization method and system based on federated learning, which solves the technical problem that the existing multi-data center joint waste heat recycling optimization method cannot balance privacy and global waste heat recycling optimization.

[0007] (II) Technical solutions

[0008] To achieve the above object, the present application is realized by the following technical solutions:

[0009] In a first aspect, the present application provides a multi-data center waste heat recovery optimization method based on federated learning, comprising:

[0010] Obtaining real-time multi-source data;

[0011] Performing feature extraction and modeling on the multi-source data through a federated learning framework, training a spatio-temporal neural network prediction model under the premise of ensuring the privacy and security of data in each data center, and obtaining future waste heat supply of each data center and thermal load demand of surrounding buildings;

[0012] Based on a waste heat recovery optimization model, performing system modeling and matching optimization on the future waste heat supply of each data center and the thermal load demand of surrounding buildings to obtain an optimal waste heat distribution strategy.

[0013] Preferably, the federated learning framework comprises a local computing layer, a secure aggregation layer and a global optimization layer;

[0014] In the local computing layer, the following is performed: performing privacy desensitization processing on the multi-source data to obtain multi-source desensitized data; using a feature selection and extraction algorithm on the multi-source desensitized data to identify and extract key variables closely related to thermal behavior, and constructing a thermodynamic feature set. Then, the sensitive features in the thermodynamic features are encrypted, and the encrypted thermodynamic features are input into a graph convolution network to generate spatio-temporal enhanced features that can simultaneously capture time trends and spatial coupling relationships;

[0015] In the secure aggregation layer, the following is performed: based on a secure multi-party computation protocol, the spatio-temporal enhanced thermodynamic features from different data centers are encrypted and aggregated to prevent original data leakage, and on this basis, the aggregation results are output to the global optimization layer in an encrypted form;

[0016] In the global optimization layer, the following is performed: the global thermodynamic features are restored by ciphertext decryption, and input into a spatio-temporal neural network model to realize prediction of future waste heat supply of each data center and thermal load demand of surrounding buildings.

[0017] Preferably, the privacy desensitization processing comprises:

[0018] Sensitive information in the multi-source data is shielded or replaced through data desensitization technology to ensure that key thermodynamic feature information is retained under the premise of protecting privacy;

[0019] Adding Laplacian noise to the business load data in the multi-source data to realize differential privacy.

[0020] Preferably, the local computing layer is also used to perform gradient inversion on the thermodynamic features of different data centers through a domain classifier, so as to align the feature distribution of the thermodynamic features of different data centers.

[0021] The domain classifier is a domain adversarial network, and its domain adversarial loss function during training is as follows:

[0022] Γ ADA =E i [logD(φ i )]+E i [logD(φ j )]

[0023] Where D represents the domain discriminator, and i, j represent samples from different data centers; φ i φ j This represents the thermodynamic characteristics of different data centers, and E is the expectation operator, which means the mean is calculated.

[0024] Preferably, the waste heat recovery and utilization optimization model includes an objective function and constraints, wherein the objective function is to maximize waste heat utilization, minimize energy loss during heat transfer, and achieve spatiotemporal load balance of the system, and its expression is as follows:

[0025]

[0026] in, This represents the proportion of waste heat output from data center i at time t; Indicates whether the heat transfer path from i to j is enabled at time t; Indicates the energy consumption of the heat pump; η1 represents the recoverable waste heat of data center i at time t; N represents the number of nodes; η1, η2, and η3 represent the output weights. The value score optimizes attention weights, and its calculation formula is as follows:

[0027]

[0028] Among them, W Q W K Represents a learnable query / key matrix; d k Characteristic dimensions representing thermodynamic features; φi and φj represent the thermodynamic features of different data centers; Represents the characteristic-related weight matrix; V represents the value score of data center i at time t; th The threshold representing the value score;

[0029] The constraints include:

[0030] Ensure the total heat delivered to all users is not less than their demand, avoiding under-supply:

[0031]

[0032] Residual heat temperature must reach the minimum district heating requirement, preventing low-temperature corrosion of pipelines:

[0033]

[0034] Limit the change range of the proportion of residual heat output, avoiding thermal shock damage to equipment:

[0035]

[0036] The total cost of residual heat utilization must be less than 80% of traditional coal-fired heating, ensuring economic feasibility:

[0037]

[0038] Only physically connected nodes can establish heat delivery paths:

[0039]

[0040] The heat delivery path of a single data center must not exceed its pipe network design capacity:

[0041]

[0042] The actual output heat must not exceed the maximum capacity of the equipment:

[0043]

[0044] The total heat supply of all data centers must meet the network requirements:

[0045]

[0046] where, represents the heat demand of heat user node j at time t, represents the cost of delivering one unit of heat from i to j, C coal represents the equivalent cost of coal-fired heating, represents the residual heat output temperature of data center i, represents the maximum recoverable residual heat capacity of data center i, K i represents the maximum number of heat delivery paths simultaneously connected by data center i, represents the total heat supply demand; represents whether the heat delivery path from data center i to heat user j is enabled at time t;A ij represents whether there is a direct physical pipe connection between data center i and heat user j.

[0047] Preferably, the waste heat recycling optimization model further comprises: updating the output weight through the PPO algorithm, specifically as follows:

[0048] The state space comprises the waste heat amount, temperature, heat transfer cost, and regional heat demand of each data center, and the weight selection in the past H period; the action space is the output weight adjustment amount Δη, which is converted into a legal weight through Softmax:

[0049]

[0050] The reward function is:

[0051]

[0052] wherein, aligning with the original target, representing a penalty for weight mutation, ensuring policy smoothness; then, the CLIP target and the value target are calculated by calculating the GAE advantage, and the policy network update and the value network update are completed, and the above multi-objective optimization output weight is obtained by repeating until convergence.

[0053] Preferably, the multi-data center waste heat recycling optimization method further comprises: issuing the optimal waste heat distribution strategy to each data center and the heat supply network for execution, and monitoring the execution in real time, and dynamically adjusting the distribution scheme according to the actual situation.

[0054] In a second aspect, the present application provides a multi-data center waste heat recycling optimization system based on federated learning, comprising:

[0055] A data acquisition module is configured to acquire real-time multi-source data.

[0056] A federated learning module is configured to perform feature extraction and modeling on the multi-source data through a federated learning framework, train a spatio-temporal neural network prediction model under the premise of ensuring the data privacy and security of each data center, and obtain future waste heat supply of each data center and heat load demand of surrounding buildings.

[0057] A recycling optimization module is configured to perform system modeling and matching optimization on the future waste heat supply of each data center and the heat load demand of surrounding buildings through a waste heat recycling optimization model, so as to obtain an optimal waste heat distribution strategy.

[0058] In a third aspect, the present application provides a computer readable storage medium storing a computer program for multi-data center waste heat recycling optimization based on federated learning, wherein the computer program enables a computer to execute the multi-data center waste heat recycling optimization method based on federated learning as described above.

[0059] In a fourth aspect, the present application provides an electronic device comprising:

[0060] one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs comprise a program for executing the federated learning-based multi-data center waste heat recycling optimization method as described above.

[0061] (III) Advantages

[0062] The present application provides a federated learning-based multi-data center waste heat recycling optimization method and system. Compared with the prior art, the following advantages are achieved:

[0063] The present application extracts and models multi-source data through a federated learning framework, trains a spatio-temporal neural network prediction model under the premise of ensuring the privacy and security of data in each data center, and obtains future waste heat supply of each data center and building heat load demand in the surrounding area. Based on a waste heat recycling optimization model, the waste heat supply of each data center and the building heat load demand in the surrounding area are systematically modeled and optimized to obtain an optimal waste heat distribution strategy. The present application realizes waste heat recycling optimization in an encrypted state while ensuring that data privacy is not leaked, and breaks through the dependence on centralized data storage and the risk of privacy leakage in traditional centralized modeling. BRIEF DESCRIPTION OF DRAWINGS

[0064] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, brief descriptions will be given below of the drawings needed to be used in the embodiments or prior art descriptions. Obviously, the drawings described below are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0065] Figure 1 A block diagram of a federated learning-based multi-data center waste heat recycling optimization method according to an embodiment of the present application;

[0066] Figure 2 A flowchart of a federated learning-based multi-data center waste heat recycling optimization method according to an embodiment of the present application. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0068] The embodiment of the application provides a multi-data center waste heat recycling optimization method and system based on federated learning, solves the technical problem that the existing waste heat recycling optimization method of multiple data centers cannot consider privacy and global waste heat recycling optimization at the same time, realizes waste heat recycling optimization between multiple data centers by privacy-enhanced calculation while ensuring privacy, and realizes heat distribution between multiple data centers and surrounding area building heating networks on the premise of protecting local data privacy of each data center.

[0069] The technical solution in the embodiment of the application is to solve the above technical problems, and the general idea is as follows:

[0070] The waste heat utilization of multiple data centers mainly needs to consider the following problems: on the one hand, different data center loads fluctuate greatly, and waste heat is difficult to accurately predict and control in a single center; on the other hand, the dispatching of the surrounding area building heating network needs to consider the balance of multiple center heat sources and pipe network demand, heat loss, pipe network thermal inertia and other complex factors. More importantly, multiple data centers jointly perform waste heat recycling optimization, which needs to share their own operation data and waste heat information. However, due to business and security considerations, each data center is often unwilling to directly share original operation data or load information, because these data contain sensitive business information and there is a risk of privacy leakage, which easily forms a data island and a privacy barrier.

[0071] The existing centralized method needs to upload original data to a third-party platform, and cannot realize cross-data center recycling optimization on the premise of protecting data privacy, while the traditional distributed optimization method is difficult to consider global optimality and model accuracy.

[0072] As known from the above description, the existing waste heat recycling optimization method of multiple data centers mainly has the following defects:

[0073] (1) The existing waste heat recycling optimization method of data centers fails to fully consider the differences in waste heat generation rules and characteristics of different data centers.

[0074] (2) The existing waste heat recycling optimization method of data centers fails to consider the data privacy and security protection problem when multiple data centers are coordinated and optimized.

[0075] (3) The existing waste heat recycling optimization method of data centers only supports simple scalar data aggregation when applying the federated learning framework, and cannot effectively handle the joint optimization of multi-data center spatio-temporal data.

[0076] In order to overcome the above defects, the embodiment of the application proposes a multi-data center waste heat recycling optimization method and system based on federated learning, aiming to coordinate the heat distribution between multiple data centers and regional heating networks while protecting the local data privacy of each data center. First, the embodiment of the application protects the data privacy security of each data center while performing global waste heat utilization optimization through data privacy desensitization technology. Second, the embodiment of the application effectively processes the joint optimization of the spatio-temporal tensor data of each data center through graph convolution network and spatio-temporal neural network technology, solving the limitations of traditional federated learning framework in the spatio-temporal data scene. Finally, the embodiment of the application introduces a value-driven attention mechanism to dynamically allocate weights for different heat source value data centers, better considering the differences in waste heat characteristics of different data centers.

[0077] In order to better understand the above technical solutions, the above technical solutions will be described in detail below in conjunction with the drawings in the specification and specific embodiments.

[0078] The embodiment of the application provides a multi-data center waste heat recycling optimization method based on federated learning, as shown in Figure 1 , comprising:

[0079] S1, acquiring real-time multi-source data;

[0080] S2, performing feature extraction and modeling on the multi-source data through a federated learning framework, training a spatio-temporal neural network prediction model under the premise of ensuring the data privacy security of each data center, and obtaining future waste heat supply of each data center and heat load demand of surrounding buildings;

[0081] S3, based on a waste heat recycling optimization model, performing system modeling and matching optimization on the future waste heat supply of each data center and the heat load demand of surrounding buildings, to obtain an optimal waste heat distribution strategy;

[0082] The federated learning framework comprises a local computing layer, a secure aggregation layer and a global optimization layer.

[0083] In the local computing layer, the following is performed: the multi-source data is subjected to privacy desensitization processing to obtain multi-source desensitization data; a feature selection and extraction algorithm is used on the multi-source desensitization data to identify and extract key variables closely related to thermal behavior, and a thermodynamic feature set is constructed. Then, the sensitive features in the thermodynamic features are encrypted, and the encrypted thermodynamic features are input into a graph convolution network to generate spatio-temporal enhanced features that can simultaneously capture time trends and spatial coupling relationships;

[0084] In the secure aggregation layer, the following is performed: the spatio-temporal enhanced thermodynamic features from different data centers are encrypted and aggregated based on a secure multi-party computation protocol, effectively preventing original data leakage. On this basis, the aggregation result is output to the global optimization layer in an encrypted form.

[0085] In the global optimization layer, the global thermodynamic features are recovered by ciphertext decryption and input into the spatio-temporal neural network model to realize the prediction of the future waste heat supply of each data center and the heat load demand of the surrounding area buildings.

[0086] The embodiment of the present application realizes the optimization of waste heat recycling in the encrypted state under the premise of ensuring that the data privacy is not leaked, and breaks through the dependence on centralized storage of data and the privacy leakage risk problem in traditional centralized modeling.

[0087] The embodiment of the present application will be described in detail below based on the specific flowchart as shown in Figure 2

[0088] In an embodiment, S1, real-time multi-source data is acquired, specifically including:

[0089] Each data center collects local operation data of each data center through intelligent devices such as sensors, including IT equipment data, i.e., server cabinet power consumption, CPU / GPU temperature, load rate. Cooling system data, i.e., cooling water flow, inlet and outlet water temperature, pump speed, etc. Environmental data, i.e., computer room temperature and humidity, external weather data. At the same time, the history data of regional heating network heat demand is collected, including time, environmental temperature, building type, etc.

[0090] In an embodiment, S2, the multi-source data is subjected to feature extraction and modeling through the pre-constructed federated learning framework, the spatio-temporal neural network prediction model is trained under the premise of ensuring the data privacy and security of each data center, and the future waste heat supply of each data center and the heat load demand of the surrounding area buildings are obtained. The specific implementation process is as follows:

[0091] The pre-constructed federated learning framework includes a local computing layer, a secure aggregation layer and a global optimization layer. The training process is as follows:

[0092] The collected data is subjected to privacy desensitization processing in the local computing layer to obtain multi-source desensitized data, the sensitive information (identity, geographic coordinates) in the multi-source data is shielded or replaced, and the numerical data (such as temperature value, flow value) is retained. Laplace noise (ε=0.1) is added to the business load data to realize differential privacy.

[0093] After obtaining the multi-source desensitized data, thermodynamic feature extraction is performed based on the multi-source desensitized data. The multi-source desensitized data is subjected to feature selection and extraction algorithm, the key variables closely related to the thermal behavior are identified and extracted, and the original heterogeneous data (such as server power consumption, cooling water flow, environmental temperature and humidity) is converted into standardized thermodynamic features:

[0094] φ​i = [η i , ΔT i , τ i , α i ]

[0095] wherein, η i represents the waste heat recovery efficiency coefficient, ΔT i represents the temperature difference between waste heat and environment, τ i represents the thermal response time constant, and α i represents the heat flow fluctuation rate.

[0096] At the same time, the domain classifier is embedded in the local computing layer, and the feature distribution of the thermodynamic characteristics of different data centers is aligned through gradient inversion to map it to a unified feature space. The domain classifier is a domain adversarial network, and the domain adversarial loss function in the training process is as follows:

[0097] Γ ADA = E i [logD(φ i )] + E j [logD(φ j )]

[0098] wherein, D represents the domain discriminator, i and j represent samples of different data centers; φ i and φ j represent thermodynamic characteristics of different data centers, E is an expectation operator, and represents mean calculation.

[0099] The Paillier encryption algorithm is used for partial homomorphic encryption of the thermodynamic characteristics: by analyzing the thermodynamic characteristics, the sensitive key features are encrypted by Paillier, and the non-key features (auxiliary environmental parameters) are processed in plaintext.

[0100] The encrypted thermodynamic characteristics are sent to each data center through the developed communication protocol (HTTPS). Each data center receives the global model parameters, reverses the sequence of the model parameters and loads them into the local model. The historical multi-source desensitization data are loaded from the local computing layer, the training set and the verification set are divided, the model is trained using the training set and the verification set, the encrypted thermodynamic characteristics are processed using the graph convolution network to retain the spatio-temporal dependence. In the training, the Dropout layer is used to prevent overfitting, the adaptive learning rate is adjusted, and the early stopping mechanism is supported to avoid overtraining.

[0101] In the security aggregation layer, hierarchical differential privacy is performed, a spatio-temporal adaptive strategy and a federated average algorithm are used to aggregate parameters, and the specific implementation process is as follows:

[0102] Layered differential privacy protection is performed. When adding Laplace noise (ε≤0.5) during local model training, relatively low noise (but still satisfying the ε constraint) is injected into the gradient update of important features (such as high-weight thermodynamic parameters), and slightly higher noise is injected into secondary features. At the same time, a space-time adaptive strategy is adopted to dynamically adjust the encryption range and noise intensity according to the local model training stage and space-time data characteristics. During the simulation of secure parameter aggregation, the oblivious transfer (OT) technology is used to realize secure aggregation. After training is completed, each data center uploads the processed local model parameters to the aggregator, applies zero-knowledge proof (ZKP) to verify the compliance of the data uploaded by each node, and avoids malicious nodes from injecting false data. The aggregator uses the federated averaging algorithm to aggregate the parameters, uses a more simplified aggregation method for non-critical parameters in the early training rounds to reduce the encryption calculation amount, and strictly encrypts and aggregates the critical parameters close to convergence to balance privacy and precision. The secure aggregation layer outputs encrypted global features to the global optimization layer.

[0103] The global optimization layer decrypts the encrypted global features and inputs them into the space-time neural network model for updating and evaluation. After each iteration, the noise allocation and encryption range are adjusted according to the validation set results until the model converges, obtaining the global model.

[0104] The trained federated learning framework processes real-time multi-source data as follows:

[0105] The following operations are performed in the local computing layer:

[0106] The multi-source data is subjected to privacy desensitization processing to obtain multi-source desensitized data. Feature selection and extraction algorithms are used on the multi-source desensitized data to identify and extract key variables closely related to thermal behavior, and a thermodynamic feature set is constructed. The sensitive features in the thermodynamic features are subjected to Paillier homomorphic encryption, and the encrypted features are input into a graph convolution network to generate space-time enhanced features that can capture both time trends and spatial coupling relationships. The space-time enhanced features are uploaded to the secure aggregation layer;

[0107] In the secure aggregation layer, the space-time enhanced thermodynamic features from different data centers are encrypted and aggregated based on secure multi-party computation protocols, effectively preventing the leakage of raw data. On this basis, the aggregation results are output in encrypted form to the global optimization layer;

[0108] In the global optimization layer, the global thermodynamic features are restored through ciphertext decryption, and they are input into the space-time neural network model to realize the prediction of future waste heat supply of each data center and building heat load demand of the surrounding area.

[0109] In an embodiment, S3 adopts a waste heat recycling optimization model to model and optimize the system of waste heat supply of each data center in the future and the heat load demand of surrounding buildings, and obtain an optimal waste heat distribution strategy. Specifically, the method comprises the following steps:

[0110] In the embodiment of the present application, an attention mechanism is introduced into the waste heat recycling optimization model, and a characteristic correlation weight matrix is designed.

[0111]

[0112] wherein W Q , W K represent the learnable query / key matrix, d k represents the feature dimension.

[0113] High-value heat source determination is performed, and a value score V i t The attention weight is optimized based on the value score.

[0114]

[0115] By introducing the attention weight, the heterogeneity of different data centers can be distinguished, so that the high-value nodes have a greater impact on the objective function, and the optimization direction can be indirectly guided by scaling the target item. V th represents the value score of the data center i at time t; V represents the threshold value of the value score; when the value score of a certain data center exceeds the threshold value V th , the attention weight thereof will be scaled up. This formula is only applied when , otherwise the original weight is kept.

[0116] The specific waste heat recycling optimization model includes an objective function and constraint conditions, and the objective function is to maximize waste heat utilization, minimize energy loss in the heat transfer process, and balance the space-time of system load. The expression is as follows:

[0117]

[0118] wherein, represents the waste heat output ratio of the data center i at time t, represents whether the heat transfer path from i to j is enabled at time t. represents the heat pump energy consumption, represents the recoverable waste heat amount of the data center i at time t, and N represents the number of nodes. η1, η2, and η3 represent output weights.

[0119] It should be noted that in the embodiments of the present application, considering that the waste heat supply and demand is dynamically changing, the weight needs to be adjusted in real time in a dynamic environment, and here the PPO algorithm is applied to update the weight strategy, the state space includes the waste heat amount, temperature, heat transfer cost, regional heat demand, and past H period weight selection of each data center. The action space is the output weight adjustment amount Δη, which is converted into a legal weight by Softmax:

[0120]

[0121] The reward function is:

[0122]

[0123] Wherein, Align with the original target, Indicates the punishment of weight mutation, and ensures the smoothness of the strategy. Then, by calculating the GAE advantage, the CLIP target and the value target are calculated, the policy network update and the value network update are completed, and the repetition is repeated until convergence, so as to obtain the above multi-objective optimization weight. That is, the output weights η1, η2 and η3 are updated by the PPO algorithm.

[0124] The constraint conditions include:

[0125] Ensure that the total heat delivered to all users is not lower than their demand, and avoid insufficient heating:

[0126]

[0127] The waste heat temperature must reach the minimum heating requirement of the region to prevent low-temperature corrosion of the pipeline:

[0128]

[0129] Limit the change range of the waste heat output ratio to avoid damage to equipment due to heat shock:

[0130]

[0131] The total cost of waste heat utilization needs to be lower than 80% of the traditional coal-fired heating to ensure economic feasibility:

[0132]

[0133] Only physically connected nodes can establish a heat transfer path:

[0134]

[0135] The heat transfer path of a single data center should not exceed its pipe network design capacity:

[0136]

[0137] The actual heat output must not exceed the equipment's maximum capacity:

[0138]

[0139] The total heat supply of all data centers must meet the network's requirements:

[0140]

[0141] in, This represents the hot demand of hot user node j at time t. C represents the cost of transferring a unit of heat from i to j. coal This represents the equivalent cost of coal-fired heating. This indicates the waste heat output temperature of data center i. K represents the maximum recoverable waste heat capacity of data center i. i This represents the maximum number of heat transfer paths that data center i can connect to simultaneously. This indicates the total heating demand. Indicates whether the heat transfer path from node i (data center) to node j (heat user) is enabled at time t. This is a binary decision variable; 1 indicates enabled, 0 indicates disabled. ij This indicates whether there is a direct physical pipe connection between node i and node j; 1 indicates a connection, and 0 indicates no connection.

[0142] The data on the future waste heat generation of each data center and the heat demand of the regional heating system in the future period are input into the waste heat recovery and utilization optimization model to obtain the optimal waste heat allocation strategy. The optimal waste heat allocation strategy includes the heat delivered by each data center to the regional heating network, the delivery time, and the path.

[0143] In one embodiment, the multi-data center waste heat recovery and utilization optimization method based on federated learning further includes: S4, distributing the optimal waste heat allocation strategy to each data center and heating network for execution, monitoring the execution status in real time, and dynamically adjusting the allocation scheme according to the actual situation.

[0144] This invention also provides a multi-datacenter waste heat recovery and utilization optimization system based on federated learning, comprising:

[0145] The data acquisition module is used to acquire real-time multi-source data;

[0146] The federated learning module is used to extract features and model multi-source data through the federated learning framework. Under the premise of ensuring the data privacy and security of each data center, it trains a spatiotemporal neural network prediction model to obtain the future waste heat supply of each data center and the heat load demand of surrounding buildings.

[0147] A recycling optimization module is configured to perform system modeling and matching optimization on future waste heat supply of each data center and heat load demand of surrounding area buildings by a waste heat recycling optimization model, so as to obtain an optimal waste heat distribution strategy.

[0148] The federated learning framework comprises a local computing layer, a secure aggregation layer and a global optimization layer.

[0149] In the local computing layer, the following is performed: privacy desensitization processing is performed on the multi-source data to obtain multi-source desensitized data; a feature selection and extraction algorithm is used on the multi-source desensitized data to identify and extract key variables closely related to the thermal behavior, thereby constructing a thermodynamic feature set. Then, sensitive features in the thermodynamic features are encrypted, and the encrypted thermodynamic features are input into a graph convolution network to generate spatiotemporal enhanced features capable of capturing time trends and spatial coupling relationships simultaneously.

[0150] In the secure aggregation layer, the following is performed: the spatiotemporal enhanced thermodynamic features from different data centers are encrypted and aggregated based on a secure multi-party computation protocol, thereby effectively preventing original data leakage. On this basis, the aggregation results are output to the global optimization layer in an encrypted form.

[0151] In the global optimization layer, the following is performed: the global thermodynamic features are recovered by ciphertext decryption, and are input into a spatiotemporal neural network model, thereby realizing prediction of future waste heat supply of each data center and heat load demand of surrounding area buildings.

[0152] In the specific embodiment process, a performing and feedback module is further included, which is configured to perform the optimal waste heat distribution strategy to each data center and the heating network, and to monitor the execution in real time, and to dynamically adjust the distribution scheme according to the actual situation.

[0153] It can be understood that the multi-data center waste heat recycling optimization system based on federated learning provided by the embodiment of the present application corresponds to the multi-data center waste heat recycling optimization method based on federated learning described above, and the explanation, examples, beneficial effects and other parts of the related content can refer to the corresponding content in the multi-data center waste heat recycling optimization method based on federated learning, which will not be repeated here.

[0154] The embodiment of the present application further provides a computer readable storage medium storing a computer program for multi-data center waste heat recycling optimization based on federated learning, wherein the computer program causes a computer to execute the multi-data center waste heat recycling optimization method based on federated learning as described above.

[0155] The embodiment of the present application also provides an electronic device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs comprise a program for executing the federated learning-based multi-data center waste heat recycling optimization method as described above.

[0156] In summary, compared with the prior art, the following beneficial effects are achieved:

[0157] 1. The embodiment of the present application realizes the optimization of waste heat recycling of each data center in an encrypted state under the premise of ensuring that the data privacy of each data center is not leaked, and obtains a global optimal waste heat distribution strategy.

[0158] 2. The embodiment of the present application effectively captures the spatio-temporal correlation of waste heat distribution based on the graph convolution network and the spatio-temporal neural network technology, and solves the problems of slow model convergence and low precision of traditional federated learning in the spatio-temporal coupling scene.

[0159] 3. The embodiment of the present application eliminates the differences in waste heat characteristics across data centers through the value-driven dynamic attention mechanism, improves the waste heat utilization rate of different data centers, and reduces the supply-demand matching error of the overall heating network.

[0160] It should be noted that, in this document, relational terms such as first and second and the like can merely be used to distinguish one entity or action from another, without necessarily requiring or implying any actual such relationship or order between or among the entities or actions. Also, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without limitation, an element preceded by "comprises... a" does not, without more constraints, foreclose the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0161] The above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An optimization method for waste heat recovery and utilization in multiple data centers based on federated learning, characterized in that, include: Acquire real-time multi-source data; By using a federated learning framework to extract features and model multi-source data, and training a spatiotemporal neural network prediction model while ensuring the data privacy and security of each data center, the future waste heat supply of each data center and the heat load demand of surrounding buildings can be obtained. Based on the waste heat recovery and utilization optimization model, the system modeling and matching optimization of the waste heat supply of each data center and the heat load demand of surrounding buildings are carried out to obtain the optimal waste heat distribution strategy.

2. The optimized method for waste heat recovery and utilization in multiple data centers as described in claim 1, characterized in that, The federated learning framework includes a local computation layer, a secure aggregation layer, and a global optimization layer; In the local computation layer, the following steps are performed: privacy-de-identifying the multi-source data to obtain de-identified multi-source data; feature selection and extraction algorithms are then used on the de-identified multi-source data to identify and extract key variables closely related to thermodynamic behavior, constructing a thermodynamic feature set. Subsequently, sensitive features within the thermodynamic features are encrypted, and the encrypted thermodynamic features are input into a graph convolutional network to generate spatiotemporal enhanced features that can simultaneously capture temporal trends and spatial coupling relationships. In the secure aggregation layer: based on the secure multi-party computation protocol, spatiotemporal enhanced thermodynamic features from different data centers are encrypted and aggregated to prevent the leakage of the original data. On this basis, the aggregation results are output to the global optimization layer in encrypted form. Execution in the global optimization layer: The global thermodynamic characteristics are recovered by decrypting the ciphertext and input into the spatiotemporal neural network model to predict the future waste heat supply of each data center and the heat load demand of surrounding buildings.

3. The optimized method for waste heat recovery and utilization in multiple data centers as described in claim 1, characterized in that, The privacy desensitization process includes: Data anonymization techniques are used to mask or replace sensitive information in multi-source data to ensure that key thermodynamic characteristics are preserved while protecting privacy. Differential privacy is achieved by adding Laplace noise to the business load data in multi-source data.

4. The optimized method for waste heat recovery and utilization in multiple data centers as described in any one of claims 1 to 3, characterized in that, The local computing layer is also used to perform gradient inversion on the thermodynamic features of different data centers through a domain classifier, so as to align the feature distribution of the thermodynamic features of different data centers. The domain classifier is a domain adversarial network, and its domain adversarial loss function during training is as follows: C ADA =E i [logD(φ i )]+E j [logD(φ j )] Where D represents the domain discriminator, and i, j represent samples from different data centers; φ i φ j This represents the thermodynamic characteristics of different data centers, and E is the expectation operator, which means the mean is calculated.

5. The optimized method for waste heat recovery and utilization in multiple data centers as described in any one of claims 1 to 3, characterized in that, The waste heat recovery and utilization optimization model includes an objective function and constraints. The objective function aims to maximize waste heat utilization, minimize energy loss during heat transfer, and achieve spatiotemporal load balance of the system. Its expression is as follows: in, This represents the proportion of waste heat output from data center i at time t; Indicates whether the heat transfer path from i to j is enabled at time t; Indicates the energy consumption of the heat pump; η1 represents the recoverable waste heat of data center i at time t; N represents the number of nodes; η1, η2, and η3 represent the output weights. The value score optimizes attention weights, and its calculation formula is as follows: Among them, W Q W K Represents a learnable query / key matrix; d k Characteristic dimensions representing thermodynamic features; φi and φj represent the thermodynamic features of different data centers; Represents the characteristic-related weight matrix; V represents the value score of data center i at time t; th The threshold representing the value score; The constraints include: Ensure that the total heat delivered to all users is not less than their needs, and avoid insufficient heating: The waste heat temperature must meet the minimum heating requirements of the area to prevent low-temperature corrosion of the pipes. Limit the variation in the waste heat output ratio to avoid thermal shock damage to the equipment. The total cost of waste heat utilization must be less than 80% of that of traditional coal-fired heating to ensure economic feasibility. Only physically connected nodes can establish a heat transfer path: The heat transfer path of a single data center must not exceed its piping network design capacity: The actual heat output must not exceed the equipment's maximum capacity: The total heat supply of all data centers must meet the network's requirements: in, This represents the hot demand of hot user node j at time t. C represents the cost of transferring a unit of heat from i to j. coal This represents the equivalent cost of coal-fired heating. This indicates the waste heat output temperature of data center i. K represents the maximum recoverable waste heat capacity of data center i. i This represents the maximum number of heat transfer paths that data center i can connect to simultaneously. Indicates total heating demand; Indicates whether the heat transfer path from data center i to hot user j is enabled at time t; A ij This indicates whether there is a direct physical pipe connection between data center i and hot user j.

6. The optimized method for waste heat recovery and utilization in multiple data centers as described in claim 5, characterized in that, The waste heat recovery and utilization optimization model also includes updating the output weights using the PPO algorithm, as follows: The state space includes the residual heat, temperature, heat transfer cost, and regional heat demand of each data center, as well as the weight selection for the past H time period; the action space is the output weight adjustment amount Δη, which is converted into valid weights via Softmax. The reward function is: in, Align with the original target. This indicates that the penalty weights change abruptly to ensure policy smoothness. Then, by calculating the GAE advantage, the CLIP objective and the value objective are calculated, and the policy network and value network are updated. This process is repeated until convergence, thus obtaining the output weights of the above multi-objective optimization.

7. The optimized method for waste heat recovery and utilization in multiple data centers as described in any one of claims 1 to 3, characterized in that, The optimization method for waste heat recovery and utilization in multiple data centers also includes: distributing the optimal waste heat allocation strategy to each data center and heating network for execution, monitoring the execution status in real time, and dynamically adjusting the allocation scheme according to the actual situation.

8. A multi-datacenter waste heat recovery and utilization optimization system based on federated learning, characterized in that, include: The data acquisition module is used to acquire real-time multi-source data; The federated learning module is used to extract features and model multi-source data through the federated learning framework. Under the premise of ensuring the data privacy and security of each data center, it trains a spatiotemporal neural network prediction model to obtain the future waste heat supply of each data center and the heat load demand of surrounding buildings. The waste heat recovery and utilization optimization module is used to systematically model and match the future waste heat supply of each data center and the heat load demand of surrounding buildings through a waste heat recovery and utilization optimization model, so as to obtain the optimal waste heat distribution strategy.

9. A computer-readable storage medium, characterized in that, It stores a computer program for optimizing the recovery and utilization of waste heat from multiple data centers based on federated learning, wherein the computer program causes a computer to execute the optimization method for the recovery and utilization of waste heat from multiple data centers based on federated learning as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the federated learning-based multi-datacenter waste heat recovery optimization method as described in any one of claims 1 to 7.

Images