Federal learning prediction method and system for demand response potential of interconnected data center

Abstract

The invention discloses a federated learning prediction method and system for the demand response potential of an interconnected data center. According to the method, historical load data of each data center is collected, so that the demand response potential is estimated; constructing a demand response potential time sequence prediction model in each data center; a data center is selected as a central server, a federated learning algorithm is adopted to train a prediction model of each data center, each non-central data center uploads iterated parameters to the central server to calculate global parameters after each round of iteration, and the global parameters are issued to each data center for the next round of iteration. Ending training until the prediction precision reaches a set precision threshold value or the number of iterations reaches a set number threshold value; and each data center predicts a demand response potential value by using the trained model based on the demand response potential estimation value. The method has the prominent advantages that the requirement for computing platform configuration is low, popularization is easy, the prediction effect is good, the flexibility of all data centers is high, efficiency is high, and the privacy disclosure risk is low.

Description

Technical Field

[0001] The present invention relates to the technical field of power demand response of interconnected data centers, and is particularly suitable for predicting and analyzing the power demand response potential of interconnected data centers, as well as optimizing their power usage efficiency and resource allocation. Background Art

[0002] In the context of the rapid increase of renewable energy around the world, it is challenging to maintain a balance between supply and demand in the face of uncertainty in the output of new energy. Traditional load tracking strategies are difficult to cope with such dynamic changes, so it is necessary to implement demand response plans to coordinate energy and load. Demand response not only enhances the flexibility of the power system during peak loads and reduces electricity demand, but also reduces operating costs and risks, and promotes the large-scale integration of renewable energy.

[0003] In traditional demand response potential prediction methods, all data from multiple data centers are usually concentrated on a single server for training and prediction. Although this centralized processing method simplifies data management and computing processes, it also brings significant risks of privacy leakage: when a large amount of data is stored on a single server or storage device, once these devices are hacked or abused by insiders, large-scale privacy data leakage may occur. In addition, during the data transmission process, the data is at risk of being intercepted, tampered with, or stolen when it is transmitted over the network. Data processing concentrated on a single server is also limited by computing power, especially when processing massive data and complex models, the performance of a single server may not be sufficient to meet the demand. At present, distributed processing has become a trend. Through the collaborative work of multiple servers, computing efficiency and system reliability can be significantly improved. However, the centralized processing method of traditional demand response potential prediction methods fails to fully utilize the advantages of distributed processing and cannot meet the requirements of modern applications for distributed computing and high availability.

[0004] CN117154722A discloses a method and system for evaluating the potential of power demand response. The patent provides a method and system for evaluating the potential of power demand response, which belongs to the technical field of power load analysis. The method includes: constructing a training set for evaluating the actual power load during the demand response invitation period, the training set for evaluating the actual power load includes: the demand response potential characteristics of multiple demand response users and the actual power load of the corresponding demand response users during the historical demand response invitation period; establishing a Gaussian process regression model, and training the Gaussian process regression model with the training set for evaluating the actual power load, the Gaussian process regression model uses a Gaussian process to describe the mapping relationship between the demand response potential characteristics and the actual power load; obtaining the demand response potential characteristics of the target user, and inputting the demand response potential characteristics of the target user into the trained Gaussian process regression model to obtain the actual power load prediction result of the target user during the demand response invitation period, and the actual power load prediction result is used as the potential evaluation result. Since this patent uses the Gaussian process regression method, the computational complexity of this patent is relatively high when processing large-scale data, and this patent requires data processing on a single server, so the configuration requirements of the computing platform for training the model are relatively high and it is difficult to popularize; in addition, the prediction results of the Gaussian process regression method are poorly interpretable, and it is difficult to intuitively explain the demand response potential, resulting in poor model prediction results. Summary of the invention

[0005] In order to solve the shortcomings of the prior art, such as high requirements for the configuration of computing platforms, difficulty in popularization, and poor model prediction effect, the present invention provides a federated learning prediction method and system for the demand response potential of interconnected data centers.

[0006] The present invention adopts the following technical solution.

[0007] The present invention discloses, on one hand, a federated learning prediction method for the demand response potential of interconnected data centers, comprising:

[0008] Step 1: Collect historical load data of each data center and use the sliding window method to estimate the demand response potential of each data center;

[0009] Step 2: Build a demand response potential time series prediction model based on deep neural network in each data center;

[0010] Step 3: Select one of the data centers as the central server and the other data centers as non-central data centers; use the federated learning algorithm to train the prediction model of each data center. After each round of iteration, each non-central data center uploads the updated parameters of the current iteration to the central server;

[0011] Step 4: The central server calculates the global parameters based on the parameters uploaded by each data center and sends them to each data center. Each data center performs the next round of iterative training based on the sent global parameters until the prediction accuracy reaches the set accuracy threshold or the number of iterations reaches the set number threshold, and the training ends;

[0012] Step 5: Based on the estimated value of demand response potential, each data center predicts the demand response potential value of each data center through the trained demand response potential time series prediction model.

[0013] More preferably,

[0014] In step 1, the demand response potential of each data center at time point t is determined by quantifying the difference between the average load and the actual load in the sliding window at time t. The expression is defined as follows:

[0015]

[0016] Among them, DR c (t) is the estimated demand response potential of the cth data center at time t; is the average load of the cth data center in the sliding window at time t, is the first sample value of the cth data center in the sliding window at time t; M c (t) is the elasticity coefficient of the cth data center at time t.

[0017] More preferably,

[0018] The elasticity coefficient of the cth data center at time t is calculated using the following formula:

[0019]

[0020] Where N is the size of the sliding window; It is the i-th sampling value in the sliding window of the c-th data center at time t; i is an integer, and i∈[1,N].

[0021] More preferably,

[0022] In step 3, when each data center trains the demand response potential time series prediction model, the local demand response potential estimation results of one or more previous time periods are used as the input of the demand response potential time series prediction model, and the local demand response potential estimation results of the current time period are used as the output of the demand response potential time series prediction model.

[0023] More preferably,

[0024] In step 4, the central server calculates the global parameters based on the parameters uploaded by each data center according to the following formula:

[0025]

[0026] in, is the jth group of parameters in the global parameters, j is an integer, and j∈[1,U], U is the total number of groups containing parameters in the global parameters; m is an integer, and m∈[1,M], M is the number of non-central data centers, is the jth set of parameters of the mth non-central data center, is the spatial similarity weight of the parameter in the mth non-central data center.

[0027] More preferably,

[0028] The calculation formula of the spatial similarity weight of the parameter in the mth non-central data center is as follows:

[0029]

[0030] where f m is the total centroid distance from the mth non-central data center to the central server, and its calculation formula is as follows:

[0031]

[0032] in, is the hth parameter value in the jth group of parameters in the mth non-central data center, h is an integer, and h∈[1,k], k is the number of parameter values ​​contained in the jth group of parameters; CE j is the centroid of the jth set of parameters in the central server.

[0033] More preferably,

[0034] The centroid of the jth group of parameters in the central server is calculated as follows:

[0035]

[0036] in is the hth parameter value in the jth set of parameters in the central server.

[0037] The present application also claims protection for a federated learning prediction system for demand response potential using the aforementioned federated learning prediction method for demand response potential, including a demand response potential estimation module, a demand response potential time series prediction model construction module, a demand response potential time series prediction model training module, and a real-time demand response potential time series prediction module:

[0038] The demand response potential estimation module collects historical load data of each data center and estimates the demand response potential of each data center using a sliding window method;

[0039] The demand response potential time series prediction model construction module constructs a demand response potential time series prediction model based on a deep neural network in each data center;

[0040] The demand response potential time series prediction model training module selects any data center as the central server and other data centers as non-central data centers; the prediction model of each data center is trained using a federated learning algorithm, and after each round of iteration, each non-central data center uploads the updated parameters of the current iteration to the central server;

[0041] The central server calculates the global parameters based on the parameters uploaded by each data center and sends them to each data center. Each data center performs the next round of iterative training based on the sent global parameters until the prediction accuracy reaches the set accuracy threshold or the number of iterations reaches the set number threshold, and the training ends;

[0042] The real-time demand response potential time series prediction module predicts the demand response potential value of each data center based on the estimated value of the demand response potential of each data center through a trained demand response potential time series prediction model.

[0043] More preferably,

[0044] In the demand response potential estimation module, the demand response potential of each data center at time point t is determined by quantifying the difference between the average load and the actual load in the sliding window at time t. The expression is defined as follows:

[0045]

[0046] Among them, DR c (t) is the estimated demand response potential of the cth data center at time t; is the average load of the cth data center in the sliding window at time t, is the first sample value of the cth data center in the sliding window at time t; M c (t) is the elasticity coefficient of the cth data center at time t.

[0047] More preferably,

[0048] The elasticity coefficient of the cth data center at time t is calculated using the following formula:

[0049]

[0050] Where N is the size of the sliding window; It is the i-th sampling value in the sliding window of the c-th data center at time t; i is an integer, and i∈[1,N].

[0051] More preferably,

[0052] In the demand response potential time series prediction model training module, when each data center trains the demand response potential time series prediction model, the local demand response potential estimation results of one or more previous time periods are used as the input of the demand response potential time series prediction model, and the local demand response potential estimation results of the current time period are used as the output of the demand response potential time series prediction model.

[0053] More preferably,

[0054] In the demand response potential time series prediction model training module, the central server calculates the global parameters based on the parameters uploaded by each data center according to the following formula:

[0055]

[0056] in, is the jth group of parameters in the global parameters, j is an integer, and j∈[1,U], U is the total number of groups containing parameters in the global parameters; m is an integer, and m∈[1,M], M is the number of non-central data centers, is the jth set of parameters of the mth non-central data center, is the spatial similarity weight of the parameter in the mth non-central data center, and its calculation formula is as follows:

[0057]

[0058] where f m is the total centroid distance from the mth non-central data center to the central server, and its calculation formula is as follows:

[0059]

[0060] in, is the hth parameter value in the jth group of parameters in the mth non-central data center, h is an integer, and h∈[1,k], k is the number of parameter values ​​contained in the jth group of parameters; CE j is the centroid of the jth group of parameters in the central server, and its formula is calculated as follows:

[0061]

[0062] in is the hth parameter value in the jth set of parameters in the central server.

[0063] On the other hand, the present application discloses an electronic device, including a processor and a storage medium; the characteristics are:

[0064] The storage medium is used to store instructions;

[0065] The processor is configured to operate according to the instructions to execute a federated learning prediction method based on the aforementioned demand response potential.

[0066] The present application also discloses a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the federated learning prediction method of the demand response potential.

[0067] The beneficial effects of the present invention are as follows:

[0068] The present invention uses a sliding window method to accurately estimate the demand response of the data center. This method helps to better understand and quantify the power demand response of the data center, thereby providing a solid foundation for formulating optimized power load transfer strategies and participating in demand response DR.

[0069] The present invention also determines the parameter weights of each data center by calculating the spatial similarity of each data center. This method fully considers the complex temporal and spatial correlation of interconnected data centers, so that the prediction model can more accurately reflect the power demand characteristics of different data centers at different geographical locations and time points.

[0070] In addition, the present invention also uses a federated learning algorithm based on a time series prediction model to realize the demand response potential prediction of interconnected data centers. Federated learning not only allows model training on local devices to maintain data privacy, but also effectively improves the accuracy and reliability of predictions through global model aggregation, and deeply considers the interconnection characteristics of interconnected data centers. Through multiple rounds of training, the federated learning algorithm significantly improves the effect of demand response potential assessment and overcomes the limitations of traditional methods in error accuracy and application effect.

[0071] Finally, the present invention also has outstanding advantages such as low configuration requirements for the computing platform, easy popularization, good prediction effect, high flexibility of each data center, high efficiency, and low risk of privacy leakage. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 is a flow chart of the federated learning prediction method of the present invention;

[0073] Figure 2 This is a MAE comparison chart of the prediction results of the demand response potential time series prediction model constructed based on LSTM in one embodiment of the present invention and the prediction results of the traditional single-point LSTM prediction model. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in this application are only part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the scope of protection of the present invention.

[0075] This application discloses a federated learning prediction method for the demand response potential of interconnected data centers, as shown in the attached Figure 1 As shown, including:

[0076] Step 1: Collect historical load data of each data center and use the sliding window method to estimate the demand response potential of each data center;

[0077] The historical load data of each data center is collected and normalized. Then, the demand response potential of each data center at time point t is determined by quantifying the difference between the average load and the actual load in the sliding window at time t. The expression is defined as follows:

[0078]

[0079] Among them, DR c (t) is the estimated demand response potential of the cth data center at time t; is the average load of the cth data center in the sliding window at time t, is the first sample value of the cth data center in the sliding window at time t; M c (t) is the elasticity coefficient of the c-th data center at time t; time t is the time at which the t-th sampling point of the collected historical load data is located; and the sliding window at time t is a sliding window starting from the t-th sampling point in the historical load data.

[0080] The elasticity coefficient of the cth data center at time t is calculated as follows:

[0081]

[0082] Where N is the size of the sliding window, that is, the number of sampling points of historical load data that can be accommodated in the sliding window; It is the i-th sampling value in the sliding window of the c-th data center at time t; i is an integer, and i∈[1,N].

[0083] Step 2: Build a demand response potential time series prediction model based on deep neural network in each data center;

[0084] The deep neural network can be a recurrent neural network RNN, a long short-term memory network LSTM, or other neural network models that can be used to predict time series.

[0085] Step 3: Select one of the data centers as the central server and the other data centers as non-central data centers; use the federated learning algorithm to train the prediction model of each data center. The model training of each data center uses the local demand response potential estimation results of one or more previous time periods as the input of the demand response potential time series prediction model, and the local demand response potential estimation results of the current time period as the output of the demand response potential time series prediction model. After each round of iteration, each non-central data center uploads the updated parameters of the current iteration to the central server; preferably, the time period is 1 day.

[0086] Step 4: The central server calculates the global parameters based on the parameters uploaded by each data center. The calculation formula of the global parameters is as follows:

[0087]

[0088] in, is the jth group of parameters in the global parameters, j is an integer, and j∈[1,U], U is the total number of groups of parameters in the global parameters or the demand response potential time series prediction model located in each data center; m is an integer, and m∈[1,M], M is the number of non-central data centers, is the jth group of parameters in the demand response potential time series prediction model located in the mth non-central data center, is the spatial similarity weight of the parameters in the demand response potential time series prediction model located in the mth non-central data center, and its calculation formula is as follows:

[0089]

[0090] where f m is the total centroid distance from the mth non-central data center to the central server, and its calculation formula is as follows:

[0091]

[0092] in, is the hth parameter value in the jth group of parameters in the demand response potential time series prediction model located in the mth non-central data center, h is an integer, and h∈[1,k], k is the number of parameter values ​​contained in the jth group of parameters; CE j is the centroid of the jth group of parameters in the demand response potential time series prediction model located in the central server;

[0093] Preferably, since the j-th group of parameters in the parameters differs depending on the model selection, the j-th group of parameters may be multidimensional. Therefore, in the preferred embodiment of the present invention, all multidimensional parameter matrices are reduced to one-dimensional matrices in the same way to determine the h-th parameter value in the j-th group of parameters in the parameters.

[0094] For example: Let the jth set of parameters of the first data center be is a two-dimensional matrix, The jth set of parameters of the mth data center is denoted as is a two-dimensional matrix of the same form,

[0095] When the jth group of parameters of the first data center is expanded to When the jth group of parameters of the mth data center is expanded in the same way:

[0096] In summary, and The parameter dimensions are the same and the expansion order is consistent. It does not matter how to expand, as long as the expansion order of each data center is consistent.

[0097] In the demand response potential time series prediction model located in the central server, the calculation formula of the centroid of the jth group of parameters in the parameters is as follows:

[0098]

[0099] in It is the hth parameter value in the jth group of parameters in the demand response potential time series prediction model located in the central server.

[0100] The global parameters calculated by the central server are sent to each data center, and each data center performs the next round of iterative training based on the sent global parameters until the preset termination condition is reached, that is, the prediction accuracy reaches the set accuracy threshold or the number of iterations reaches the set number threshold, and the training ends;

[0101] Step 5: Based on the estimated value of demand response potential, each data center predicts the demand response potential value of each data center through the trained demand response potential time series prediction model.

[0102] The present invention also discloses a federated learning prediction system for demand response potential based on the aforementioned federated learning prediction method for demand response potential, including a demand response potential estimation module, a demand response potential time series prediction model construction module, a demand response potential time series prediction model training module and a real-time demand response potential time series prediction module:

[0103] The demand response potential estimation module collects historical load data of each data center and estimates the demand response potential of each data center using a sliding window method;

[0104] By quantifying the difference between the average load and the actual load in the sliding window at time t, the demand response potential of each data center at time point t is determined. The expression is defined as follows:

[0105]

[0106] Among them, DR c (t) is the estimated demand response potential of the cth data center at time t; is the average load of the cth data center in the sliding window at time t, is the first sample value of the cth data center in the sliding window at time t; M c (t) is the elasticity coefficient of the cth data center at time t, and its calculation formula is as follows:

[0107]

[0108] Where N is the size of the sliding window; It is the i-th sampling value in the sliding window of the c-th data center at time t; i is an integer, and i∈[1,N].

[0109] The demand response potential time series prediction model construction module constructs a demand response potential time series prediction model based on a deep neural network in each data center;

[0110] The deep neural network can be a recurrent neural network RNN, a long short-term memory network LSTM, or other neural network models that can be used to predict time series.

[0111] The demand response potential time series prediction model training module selects any data center as the central server and other data centers as non-central data centers; the prediction model of each data center is trained using a federated learning algorithm, and the model training of each data center uses the local demand response potential estimation results of one or more previous time periods as the input of the demand response potential time series prediction model, and the local demand response potential estimation results of the current time period as the output of the demand response potential time series prediction model. After each round of iteration, each non-central data center uploads the updated parameters of the current iteration to the central server;

[0112] The central server calculates the global parameters based on the parameters uploaded by each data center. The calculation formula of the global parameters is as follows:

[0113] The central server calculates the global parameters based on the parameters uploaded by each data center according to the following formula:

[0114]

[0115] in, is the jth group of parameters in the global parameters, j is an integer, and j∈[1,U], U is the total number of groups of parameters in the global parameters or the demand response potential time series prediction model of each data center; m is an integer, and m∈[1,M], M is the number of non-central data centers, is the jth set of parameters of the mth non-central data center, is the spatial similarity weight of the parameter of the mth non-central data center, and its calculation formula is as follows:

[0116]

[0117] where f m is the total centroid distance from the mth non-central data center to the central server, and its calculation formula is as follows:

[0118]

[0119] in, is the hth parameter value in the jth group of parameters in the mth non-central data center, h is an integer, and h∈[1,k], k is the number of parameter values ​​contained in the jth group of parameters; CE j is the centroid of the jth group of parameters in the central server, and its formula is calculated as follows:

[0120]

[0121] in is the hth parameter value in the jth set of parameters in the central server.

[0122] The global parameters calculated by the central server are sent to each data center, and each data center performs the next round of iterative training based on the sent global parameters until the prediction accuracy reaches the set accuracy threshold or the number of iterations reaches the set number threshold, and the training ends;

[0123] The real-time demand response potential time series prediction module predicts the demand response potential value of each data center based on the estimated value of the demand response potential of each data center through a trained demand response potential time series prediction model.

[0124] Embodiment 1:

[0125] A federated learning approach to predicting the demand response potential of interconnected data centers.

[0126] Step 1: Collect the historical load data of each data center, use the sliding window method to estimate the demand response potential of each data center, and use the estimation results as the data set.

[0127] First, collect historical load data from multiple data centers. Then use the sliding window method to estimate the demand response potential of each data center under the central server. The specific process is as follows:

[0128] First, collect the load data of multiple data centers. Assume there are A data centers, collect the historical load data of each data center. Then use the following formula to normalize the historical load data of each data center. The specific formula is as follows:

[0129]

[0130] in, is the normalized result of the historical load data of the cth data center, P c is the historical load data value collected from the cth data center, μ c is the average value calculated based on all historical load data of the cth data center, σ c It is the standard deviation calculated based on all historical load data of the cth data center; the value range of c is [1, A].

[0131] The sliding window technology is used to capture time elasticity, that is, by quantifying the difference between the average load and the actual load in the sliding window at time t, the estimated value of the demand response potential of each data center at time point t is determined, and the estimated results are used as the data set.

[0132] The estimated demand response potential DR of the cth data center at time t c The calculation formula of (t) is as follows:

[0133]

[0134] in, is the average load of the cth data center in the sliding window at time t, is the first sample value of the cth data center in the sliding window at time t; M c (t) is the elasticity coefficient of the cth data center at time t.

[0135] The elasticity coefficient of the cth data center at time t is calculated using the following formula:

[0136]

[0137] Where N is the size of the sliding window; It is the i-th sampling value in the sliding window of the c-th data center at time t; i is an integer, and i∈[1,N].

[0138] Step 2: Build a demand response potential time series prediction model based on deep neural network in each data center;

[0139] Step 3: Select any data center as the central server and other data centers as non-central data centers; use the federated learning algorithm to train the prediction model of each data center.

[0140] First, model initialization and parameter distribution are performed: the central server initializes a model and sends the initial parameter matrix ω0 to all data centers based on historical experience.

[0141] Secondly, the demand response potential time series prediction model is used to realize local training and parameter update: after each data center receives the parameters, it uses the local data set to train the demand response potential time series prediction model. And in each round of iteration, the model parameters are updated according to the following formula:

[0142]

[0143] in represents the jth group of parameters of the mth non-central data center in the tth iteration, represents the jth group of parameters of the mth non-central data center in the t-1th iteration, η is the set learning rate, preferably, the learning rate can be set to 0.01, It is the value obtained by substituting the derivative of the current loss function into the current parameter.

[0144] Then, parameter transmission and collection are carried out: after each round of data interaction, the non-central data center synchronizes the locally updated model parameters to the central server. The central server is then responsible for collecting and integrating the model parameters from each data center to calculate the global parameters.

[0145] After the central server collects the updated model parameters of each data center, it calculates the global parameters based on the parameters uploaded by each non-central data center according to the following formula:

[0146]

[0147] in, is the jth group of parameters in the global parameters, j is an integer, and j∈[1,U], U is the total number of groups containing parameters in the global parameters; m is an integer, and m∈[1,M], M is the number of non-central data centers, is the jth set of parameters of the mth non-central data center, is the spatial similarity weight of the parameter in the mth non-central data center, and its calculation formula is as follows:

[0148]

[0149] where f m is the total centroid distance from the mth non-central data center to the central server, and its calculation formula is as follows:

[0150]

[0151] in, is the distance from the jth group of parameters in the mth non-central data center to the centroid of the jth group of parameters in the central server, and its calculation formula is as follows:

[0152]

[0153] in, is the hth parameter value in the jth group of parameters in the mth non-central data center, h is an integer, and h∈[1,k], k is the number of parameter values ​​contained in the jth group of parameters; CE j is the centroid of the jth group of parameters in the central server, and its calculation formula is as follows:

[0154]

[0155] in is the hth parameter value in the jth set of parameters in the central server.

[0156] Global parameter update and distribution: The central server transmits the updated global parameters to the load aggregators in each data center, and the load aggregators use these global parameters to update their local parameters.

[0157] The entire process is repeated in an iterative loop until the prediction accuracy reaches the set accuracy threshold or the number of iterations reaches the set number threshold, and the training ends.

[0158] Embodiment 2:

[0159] A federated learning approach to predicting the demand response potential of interconnected data centers.

[0160] The dataset used includes 96 load points from 10 data centers on a demand response day. These experimental data come from the data center of the National Renewable Energy Laboratory NREL-RSF in Golden, Colorado. For collaborative training, a central server and ten clients were established. Each client represents a data center.

[0161] First, the daily load of each data center is averaged to obtain the daily average load. The daily average load data is then aggregated into the monthly average load. The monthly average load data of each data center is then uniformly normalized. Finally, the processed data is used to obtain the demand response potential on the demand response day, and then divided into a training set and a test set in a ratio of 8:2.

[0162] After data processing, the LSTM model is first selected locally as the time series prediction model for training. After each round of training, the parameters are transmitted back to the central server to summarize the updated model parameters of all data centers. Then, the weights of the parameters of each data center are calculated according to the spatial similarity-based method proposed in the present invention, and the global model parameters are calculated accordingly. The global parameters are then distributed to each data center as the starting point for the next round of training. The whole process is repeated until the prediction accuracy reaches the set accuracy threshold or the number of iterations reaches the set number threshold, and the training ends.

[0163] After 5 rounds of iterations, the proposed algorithm significantly improved the evaluation effect of demand response potential. Figure 2 As shown, compared with traditional methods, the present invention performs well in terms of error accuracy, especially in reducing the mean absolute error (MAE), which is significantly better than the traditional local single training method.

[0164] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0165] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples of computer-readable storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the foregoing. As used herein, a computer-readable storage medium is not to be interpreted as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

[0166] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0167] The computer program instructions for performing the operation of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions may be executed completely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be customized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present disclosure.

[0168] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A federated learning prediction method for the demand response potential of interconnected data centers, characterized in that: include: Step 1: Collect historical load data of each data center and use the sliding window method to estimate the demand response potential of each data center; Step 2: Build a demand response potential time series prediction model based on deep neural network in each data center; Step 3: Select one of the data centers as the central server and the other data centers as non-central data centers; use the federated learning algorithm to train the prediction model of each data center. After each round of iteration, each non-central data center uploads the updated parameters of the current iteration to the central server; Step 4: The central server calculates the global parameters based on the parameters uploaded by each data center and sends them to each data center. Each data center performs the next round of iterative training based on the sent global parameters until the prediction accuracy reaches the set accuracy threshold or the number of iterations reaches the set number threshold, and the training ends; Step 5: Based on the estimated value of demand response potential, each data center predicts the demand response potential value of each data center through the trained demand response potential time series prediction model.

2. The method for predicting demand response potential through federated learning according to claim 1, characterized in that: In step 1, the demand response potential of each data center at time point t is determined by quantifying the difference between the average load and the actual load in the sliding window at time t. The expression is defined as follows: Among them, DR c (t) is the estimated demand response potential of the cth data center at time t; is the average load of the cth data center in the sliding window at time t, is the first sample value of the cth data center in the sliding window at time t; M c (t) is the elasticity coefficient of the cth data center at time t.

3. The method for predicting demand response potential through federated learning according to claim 2, characterized in that: The elasticity coefficient of the cth data center at time t is calculated using the following formula: Where N is the size of the sliding window; It is the i-th sampling value in the sliding window of the c-th data center at time t; i is an integer, and i∈[1,N].

4. The method for predicting demand response potential through federated learning according to claim 1, characterized in that: In step 3, when each data center trains the demand response potential time series prediction model, the local demand response potential estimation results of one or more previous time periods are used as the input of the demand response potential time series prediction model, and the local demand response potential estimation results of the current time period are used as the output of the demand response potential time series prediction model.

5. The method for predicting demand response potential through federated learning according to claim 1, characterized in that: In step 4, the central server calculates the global parameters based on the parameters uploaded by each data center according to the following formula: in, is the jth group of parameters in the global parameters, j is an integer, and j∈[1,U], U is the total number of groups containing parameters in the global parameters; m is an integer, and m∈[1,M], M is the number of non-central data centers, is the jth set of parameters of the mth non-central data center, is the spatial similarity weight of the parameter in the mth non-central data center.

6. The method for predicting demand response potential through federated learning according to claim 5, characterized in that: The calculation formula of the spatial similarity weight of the parameter in the mth non-central data center is as follows: where f m is the total centroid distance from the mth non-central data center to the central server, and its calculation formula is as follows: in, is the hth parameter value in the jth group of parameters in the mth non-central data center, h is an integer, and h∈[1,k], k is the number of parameter values ​​contained in the jth group of parameters; CE j is the centroid of the jth set of parameters in the central server.

7. The method for predicting demand response potential through federated learning according to claim 6, characterized in that: The centroid of the jth group of parameters in the central server is calculated as follows: in is the hth parameter value in the jth set of parameters in the central server.

8. A federated learning prediction system for demand response potential of interconnected data centers using the method for predicting demand response potential of federated learning according to any one of claims 1 to 7, characterized in that: It includes demand response potential estimation module, demand response potential time series prediction model construction module, demand response potential time series prediction model training module and real-time demand response potential time series prediction module: The demand response potential estimation module collects historical load data of each data center and estimates the demand response potential of each data center using a sliding window method; The demand response potential time series prediction model construction module constructs a demand response potential time series prediction model based on a deep neural network in each data center; The demand response potential time series prediction model training module selects any data center as the central server and other data centers as non-central data centers; the prediction model of each data center is trained using a federated learning algorithm, and after each round of iteration, each non-central data center uploads the updated parameters of the current iteration to the central server; The central server calculates the global parameters based on the parameters uploaded by each data center and sends them to each data center. Each data center performs the next round of iterative training based on the sent global parameters until the prediction accuracy reaches the set accuracy threshold or the number of iterations reaches the set number threshold, and the training ends; The real-time demand response potential time series prediction module predicts the demand response potential value of each data center based on the estimated value of the demand response potential of each data center through a trained demand response potential time series prediction model.

9. The demand response potential federated learning prediction system according to claim 8, characterized in that: In the demand response potential estimation module, the demand response potential of each data center at time point t is determined by quantifying the difference between the average load and the actual load in the sliding window at time t. The expression is defined as follows: Among them, DR c (t) is the estimated demand response potential of the cth data center at time t; is the average load of the cth data center in the sliding window at time t, is the first sample value of the cth data center in the sliding window at time t; M c (t) is the elasticity coefficient of the cth data center at time t.

10. The demand response potential federated learning prediction system according to claim 9, characterized in that: The elasticity coefficient of the cth data center at time t is calculated using the following formula: Where N is the size of the sliding window; It is the i-th sampling value in the sliding window of the c-th data center at time t; i is an integer, and i∈[1,N].

11. The demand response potential federated learning prediction system according to claim 8, characterized in that: In the demand response potential time series prediction model training module, when each data center trains the demand response potential time series prediction model, the local demand response potential estimation results of one or more previous time periods are used as the input of the demand response potential time series prediction model, and the local demand response potential estimation results of the current time period are used as the output of the demand response potential time series prediction model.

12. The demand response potential federated learning prediction system according to claim 8, characterized in that: In the demand response potential time series prediction model training module, the central server calculates the global parameters based on the parameters uploaded by each data center according to the following formula: in, is the jth group of parameters in the global parameters, j is an integer, and j∈[1,U], U is the total number of groups containing parameters in the global parameters; m is an integer, and m∈[1,M], M is the number of non-central data centers, is the jth set of parameters of the mth non-central data center, is the spatial similarity weight of the parameter in the mth non-central data center, and its calculation formula is as follows: where f m is the total centroid distance from the mth non-central data center to the central server, and its calculation formula is as follows: in, is the hth parameter value in the jth group of parameters in the mth non-central data center, h is an integer, and h∈[1,k], k is the number of parameter values ​​contained in the jth group of parameters; CE j is the centroid of the jth group of parameters in the central server, and its formula is calculated as follows: in is the hth parameter value in the jth set of parameters in the central server.

13. An electronic device, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is used to operate according to the instructions to execute the federated learning prediction method for demand response potential according to claims 1-7.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the federated learning prediction method for demand response potential described in claims 1-7 is implemented.

Images