Load prediction method, device and equipment of data center and storage medium

Abstract

The embodiment of the present application provides a load prediction method, device and equipment of a data center and a storage medium, the method comprises the following steps: obtaining target data of multiple historical moments; the target data comprises load data of the data center and multiple types of feature data related to the load data; inputting the load of the data center and the multiple types of feature data related to the load data of the multiple historical moments before a to-be-predicted moment into a trained load prediction model to obtain a data center load prediction result; the model is used for predicting the load of the data center at a future moment based on the target data of the historical moments and the predicted feature data of the future moment. The method of the embodiment of the present application improves the accuracy of the data center load prediction by exploring valuable feature data of the future moment and guiding the target data of the historical moments to predict the load according to the explored feature data.

Description

Technical Field

[0001] This invention relates to the field of data center technology, and in particular to a method, apparatus, device, and storage medium for data center load forecasting. Background Technology

[0002] With the increasing application and demand for cloud computing, the number of data centers is growing rapidly. Due to their enormous energy consumption, data centers place immense pressure on the economy and the environment. Therefore, accurate load forecasting of data centers is crucial for developing more efficient energy management strategies to participate in demand response and achieve a win-win situation for both the power grid and data centers.

[0003] In related technologies, predicting data center load for future moments solely based on historical data—that is, using only historical data on cooling systems, IT equipment, weather, etc.—results in low accuracy in data center load prediction. Summary of the Invention

[0004] To address the problems in the prior art, embodiments of the present invention provide a method, apparatus, device, and storage medium for data center load forecasting.

[0005] Specifically, the embodiments of the present invention provide the following technical solutions:

[0006] In a first aspect, embodiments of the present invention provide a data center load forecasting method, comprising:

[0007] Acquire target data from multiple historical time points; the target data includes data center load data and multiple types of feature data related to the load data;

[0008] The data center load at multiple historical moments prior to the time to be predicted, along with multiple types of feature data related to the load data, are input into the trained load prediction model to obtain the data center load prediction result. The model is used to predict the data center load at future moments based on the target data at historical moments and the predicted feature data at future moments. The load prediction model is trained based on the target data at the first moment, the target data at the second moment, and the predicted target data at the second moment.

[0009] Furthermore, the step of inputting data center load data from multiple historical moments prior to the time to be predicted, along with multiple types of feature data related to the load data, into the trained load prediction model to obtain the data center load prediction result includes:

[0010] The load data of the data center at historical moments and multiple types of feature data related to the load data are input into the future moment feature data exploration layer of the load prediction model to obtain the predicted feature data for the future moment.

[0011] The load data of the data center at historical time and multiple types of feature data related to the load data are input into the historical time feature data embedding layer of the load prediction model to obtain the target data vector at historical time.

[0012] The target data vector of the historical moment and the feature data of the predicted future moment are input into the load prediction layer of the load prediction model to obtain the load prediction result of the data center.

[0013] Furthermore, before inputting the data center load data from multiple historical moments prior to the time to be predicted, along with multiple types of feature data related to the load data, into the trained load prediction model to obtain the data center load prediction result, the process further includes:

[0014] The target data at the first moment is input into the load prediction model to obtain the predicted data center load data at the second moment and multiple types of feature data related to the load data.

[0015] The load prediction model is trained based on the target data at the second moment, the predicted data center load data at the second moment, and multiple types of feature data related to the load data, to obtain the trained load prediction model.

[0016] Further, the step of training the load prediction model based on the acquired target data at the second time moment, the predicted data center load data at the second time moment, and multiple types of feature data related to the load data to obtain the trained load prediction model includes:

[0017] The load forecasting model is trained using the following loss function:

[0018]

[0019] N represents the target data samples from N data centers; K represents the dimension of the target data; This represents the predicted value of the i-th type of feature data in the j-th target data sample at time T+1. This represents the true value of the feature data of the i-th type in the j-th target data sample at time T+1.

[0020] Furthermore, the future moment feature data exploration layer of the load prediction model is constructed based on a multi-task learning network architecture and a long short-term memory network; the historical moment feature data embedding layer of the load prediction model is constructed based on a long short-term memory network; and the load prediction layer of the prediction model is constructed based on a fully self-attention network and a multilayer perceptron.

[0021] Furthermore, before inputting the data center load data from multiple historical moments prior to the time to be predicted, along with multiple types of feature data related to the load data, into the trained load prediction model to obtain the data center load prediction result, the process further includes:

[0022] The feature data is preprocessed; the preprocessing includes: determining the maximum mutual information coefficients between each type of feature data and the load data of the data center, and deleting feature data whose maximum mutual information coefficients are less than a threshold.

[0023] Secondly, embodiments of the present invention also provide a data center load forecasting device, comprising:

[0024] The acquisition module is used to acquire target data from multiple historical moments; the target data includes data center load data and multiple types of feature data related to the load data;

[0025] The prediction module is used to input the data center load of multiple historical time points before the time to be predicted and multiple types of feature data related to the load data into the trained load prediction model to obtain the data center load prediction result; the model is used to predict the data center load at future time points based on the target data at historical time points and the predicted feature data at future time points; the load prediction model is trained based on the target data at the first time point, the target data at the second time point, and the predicted target data at the second time point.

[0026] Thirdly, embodiments of the present invention also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the data center load forecasting method as described in the first aspect.

[0027] Fourthly, embodiments of the present invention also provide a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the data center load forecasting method as described in the first aspect.

[0028] Fifthly, embodiments of the present invention also provide a computer program product, including a computer program that, when executed by a processor, implements the data center load forecasting method as described in the first aspect.

[0029] The data center load forecasting method, apparatus, device, and storage medium provided in this invention acquire data center load data and multiple types of load-related feature data from multiple historical moments, and input them into a load forecasting model to predict data center load. This enables data center load forecasting based on target data from historical moments and predicted feature data from future moments. In other words, in the process of forecasting data center load, not only is the correlation between historical feature data learned, but valuable feature data from future moments is also fully extracted to forecast data center load, making the load forecasting results more accurate. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0031] Figure 1 This is a flowchart illustrating the data center load forecasting method provided in an embodiment of the present invention;

[0032] Figure 2 This is a schematic diagram of a data center load forecasting method provided in an embodiment of the present invention;

[0033] Figure 3 This is the data center load prediction model provided in the embodiments of the present invention;

[0034] Figure 4 This is a schematic diagram of the structure of the data center load forecasting device provided in an embodiment of the present invention;

[0035] Figure 5 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0037] The method described in this invention can be applied to data center scenarios to achieve data center load prediction.

[0038] In related technologies, predicting data center load for future moments solely based on historical data—that is, using only historical data such as cooling system data, IT equipment data, and weather data—results in low accuracy of data center load predictions.

[0039] The data center load forecasting method of this invention acquires data center load data from multiple historical moments and multiple types of feature data related to the load data, and inputs them into a load forecasting model to forecast the data center load. This method achieves data center load forecasting based on target data from historical moments and predicted feature data from future moments. In other words, in the process of forecasting data center load, it not only learns the correlation between historical feature data, but also fully mines valuable feature data from future moments to forecast data center load, making the load forecasting results more accurate.

[0040] The following is combined with Figures 1-5 The technical solution of the present invention will be described in detail with reference to specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0041] Figure 1 This is a flowchart illustrating an embodiment of the data center load forecasting method provided by this invention. Figure 1 As shown, the method provided in this embodiment includes:

[0042] Step 101: Obtain target data from multiple historical time points; the target data includes data center load data and multiple types of feature data related to the load data;

[0043] Specifically, in related technologies, predicting data center load for future moments is done using historical data. This means that the prediction is based solely on historical data on features such as cooling systems, IT equipment, and weather. This approach lacks full utilization of load-related features at the time of prediction, resulting in low accuracy in data center load prediction.

[0044] To address the aforementioned issues, this embodiment of the invention first acquires load data from multiple historical moments and multiple types of feature data related to the load data. The feature data related to the load data includes factors that affect the data center load, such as the data center's air inlet temperature and cooling tower drainage volume.

[0045] Step 102: Input the data center load data from multiple historical moments prior to the time to be predicted, along with various types of load-related feature data, into the trained load prediction model to obtain the data center load prediction results. The model is used to predict the data center load at future moments based on the target data from historical moments and the predicted feature data from future moments. The load prediction model is trained based on the target data from the first moment, the target data from the second moment, and the predicted target data from the second moment.

[0046] Specifically, after obtaining load data of the data center at multiple historical moments prior to the time to be predicted, as well as multiple types of feature data related to the load data, this embodiment of the invention inputs the load data and feature data from multiple historical moments into a trained load prediction model to obtain the load prediction result of the data center. The load prediction model is used to predict the load of the data center at future moments based on the target data at historical moments and the predicted feature data at future moments. That is, in the process of predicting the load of the data center at future moments, it is not only based on the target data at historical moments, but also on the target data at historical moments and the predicted feature data at future moments. In other words, in the process of predicting the data center load, it not only learns the correlation between historical feature data, but also fully mines valuable feature data from future moments to predict the data center load, making the load prediction result more accurate.

[0047] For example, by predicting factors such as future cooling tower drainage volume, which is highly correlated with the future cooling system load, obtaining this characteristic data in advance can greatly help in predicting the overall data center load and improve its accuracy. Figure 2 As shown, in related technologies, valuable feature data x at future moments t+1 It cannot be obtained in advance, and the existing prediction model only learns historical features x1 to x2. t The correlation between them was not fully exploited to extract useful features from future moments. t+1 This limits model performance. The load forecasting model in this embodiment of the invention can capture dynamic dependencies among numerous historical features to explore valuable future feature data, and guide more accurate load forecasting based on this future feature data, thus improving the accuracy of data center load forecasting.

[0048] The method described in the above embodiments acquires data load data of the data center at multiple historical moments and multiple types of feature data related to the load data, and inputs them into the load prediction model to predict the data center load. This achieves the prediction of data center load based on target data at historical moments and feature data at predicted future moments. In other words, in the process of predicting data center load, it not only learns the correlation between historical feature data, but also fully mines valuable feature data at future moments to predict data center load, making the load prediction results more accurate.

[0049] In one embodiment, data center load data from multiple historical moments prior to the time to be predicted, along with multiple types of load-related feature data, are input into a trained load prediction model to obtain data center load prediction results, including:

[0050] The load data of the data center at historical moments and multiple types of load-related feature data are input into the future moment feature data exploration layer of the load prediction model to obtain the predicted future moment feature data.

[0051] The load data of the data center at historical moments and multiple types of load data-related feature data are input into the historical moment feature data embedding layer of the load prediction model to obtain the target data vector at historical moments.

[0052] By inputting the target data vector from historical moments and the predicted feature data from future moments into the load forecasting layer of the load forecasting model, the load forecasting results for the data center are obtained.

[0053] Specifically, in this embodiment of the invention, load data and feature data from historical moments are input into the future moment feature data exploration layer of the load prediction model to obtain the predicted future moment feature data. Thus, data center load can be predicted based on the target data from historical moments and the predicted future moment feature data, effectively improving the accuracy of load prediction.

[0054] For example, the target data X = {y1, y2, ..., y} from multiple historical moments. T The future time-moment feature data exploration layer of the input load forecasting model is used to obtain the predicted future time-moment feature data, where... in M represents the number of elements in the input target data. For example, if the number of feature data in the target data is 10, and the number of data center load data in the target data is 1, then M is 11; T represents the time step of the input sequence, which is the target data at multiple times.

[0055] To accurately predict future feature data based on input historical target data, optionally, a future feature data exploration layer of the load prediction model can be constructed using a multi-task learning network architecture and a long short-term memory network. This means that the future feature data exploration layer of the prediction model, constructed using multi-task learning, accurately predicts the feature data for future moments. The multi-task learning based on the long short-term memory network can utilize useful information from historical data to help analyze various types of future data. Optionally, the target data from T historical moments can be used to predict the feature data for future moments, i.e., using M-dimensional time series data X = {y1, y2, ..., y...} from the previous T time steps. T Predict the feature data at the next time step T+1. K represents the number of elements in the output target data. For example, if the number of feature data in the output target data is 5, and the number of data points on data center load in the target data is 1, then K is 6. In other words, based on the temporal characteristics of various factors affecting data center load, we model the interaction relationship between multiple time series, capture the dynamic dependencies between numerous historical feature data to explore valuable feature data for future moments, and then make more accurate predictions of data center load based on the explored feature data for future moments.

[0056] The method described in the above embodiment inputs historical load data and feature data into the future feature data exploration layer of the prediction model to obtain the predicted future feature data. Then, based on the historical target data and the predicted future feature data, the data center load can be predicted, effectively improving the accuracy of load prediction.

[0057] In one embodiment, before obtaining the data center load prediction result by inputting data center load data from multiple historical times prior to the time to be predicted, along with multiple types of load-related feature data, into the trained load prediction model, the method further includes:

[0058] The target data at the first moment is input into the load forecasting model to obtain the predicted data center load data at the second moment, as well as multiple types of characteristic data related to the load data.

[0059] Based on the acquired target data at the second moment, the predicted data center load data at the second moment, and multiple types of load-related feature data, the load prediction model is trained to obtain the trained load prediction model.

[0060] Specifically, in order to enable the load forecasting model to accurately predict characteristic data at future times, and to accurately predict data center load based on target data at historical times and predicted characteristic data at future times, this embodiment of the invention trains the load forecasting model based on target data at a first time, target data at a second time, and predicted target data at the second time. Optionally, the target data at the first time is input into the load forecasting model to obtain the predicted data center load data at the second time and multiple types of characteristic data related to the load data; the load forecasting model is trained based on the obtained target data at the second time, the predicted data center load data at the second time, and multiple types of characteristic data related to the load data to obtain the trained load forecasting model; wherein, the target data at the first time can be target data from multiple historical times; optionally, in the obtained... If the difference between the actual target data at the second moment and the predicted target data at the second moment is greater than a preset value, it indicates that the predicted feature data and the predicted data center load at the second moment differ significantly from the actual feature data and data center load at the second moment, and the load prediction model needs to be further trained. Alternatively, if the difference between the actual target data at the second moment and the predicted target data at the second moment is less than or equal to a preset value, it indicates that the predicted feature data and the predicted data center load at the second moment are consistent with the actual feature data and data center load at the second moment. In this case, the trained load prediction model can accurately predict the feature data at future moments and can accurately predict the data center load at future moments based on the target data at historical moments and the predicted feature data at future moments.

[0061] For example, target data from multiple historical moments can be acquired, including data center load data and various types of feature data related to the load data. The acquired historical target data is then preprocessed to obtain a training dataset, which is then input into a model constructed based on multi-task learning and long short-term memory networks for training, resulting in a load prediction model. This load prediction model can then be used to predict the data center load at future moments based on the historical target data and the predicted feature data for future moments. Optionally, the preprocessing of the acquired historical target data to obtain the training dataset includes: determining the maximum mutual information coefficients between each type of feature data and the data center load data, deleting feature data with a maximum mutual information coefficient less than a threshold; normalizing the target data; converting the sequence of normalized data into a value predicting the next time point based on previous time points; and proportionally dividing the dataset into a training set, a validation set, and a test set.

[0062] The method described in the above embodiment inputs the target data at a first moment into the load forecasting model to obtain the predicted data center load data at a second moment and multiple types of load-related feature data. Based on the obtained target data at the second moment, the predicted data center load data at the second moment, and the multiple types of load-related feature data, the load forecasting model is trained. This allows the trained load forecasting model to accurately predict the feature data at future moments and accurately predict the data center load based on the target data at historical moments and the predicted feature data at future moments, thereby improving the accuracy of data center load forecasting.

[0063] In one embodiment, the load prediction model is trained based on the acquired target data at the second time moment, the predicted data center load data at the second time moment, and multiple types of load data-related feature data, to obtain the trained load prediction model, including:

[0064] The load forecasting model was trained using the following loss function:

[0065]

[0066] N represents the target data samples from N data centers; K represents the dimension of the target data. This represents the predicted value of the i-th type of feature data in the j-th target data sample at time T+1. This represents the true value of the feature data of the i-th type in the j-th target data sample at time T+1.

[0067] Specifically, in this embodiment of the invention, a loss function is used. The load forecasting model is trained by using a loss function to evaluate the accuracy of the data characteristics and data center load predicted by the model for future time moments. Optionally, if the difference between the actual target data at the second time moment and the predicted target data at the second time moment is greater than a preset value, that is, if the predicted value of the i-th type of feature data in the j-th target data sample at time T+1 is used... Compared with the true value of the i-th type of feature data in the j-th target data sample at time T+1 If the difference between the predicted and actual data points is greater than a preset value, it indicates that the predicted feature data and data center load at the second time step differ significantly from the actual feature data and data center load at the second time step, requiring further training of the load prediction model. Alternatively, if the difference between the actual target data and the predicted target data at the second time step is less than or equal to a preset value, that is, the predicted value of the i-th type of feature data in the j-th target data sample at time T+1... Compared with the true value of the i-th type of feature data in the j-th target data sample at time T+1 If the difference between the two is less than or equal to the preset value, it indicates that the predicted feature data and the predicted data center load at the second moment are consistent with the actual feature data and data center load at the second moment. In this case, the load prediction model trained based on the loss function can accurately predict the feature data at future moments, and can accurately predict the data center load at future moments based on the target data at historical moments and the predicted feature data at future moments.

[0068] The method described in the above embodiments evaluates the accuracy of the feature data and data center load predicted by the load model at future times using a loss function, and trains the load prediction model based on the evaluation results. This enables the trained load prediction model to accurately predict the feature data and data center load at future times, thereby improving the accuracy of data center load prediction.

[0069] In one embodiment, the future moment feature data exploration layer of the load prediction model is constructed based on a multi-task learning network architecture and a long short-term memory network; the historical moment feature data embedding layer of the load prediction model is constructed based on a long short-term memory network; and the load prediction layer of the prediction model is constructed based on a fully self-attention network and a multilayer perceptron.

[0070] Specifically, to accurately predict future feature data based on input historical target data, optionally, the future feature data exploration layer of the load forecasting model is constructed based on a multi-task learning network architecture and a long short-term memory network layer. That is, the future feature data exploration layer of the forecasting model is constructed using a multi-task learning architecture based on a long short-term memory network, thereby accurately predicting the feature data of future moments. Here, multi-task learning can utilize useful information from historical data to help learn multiple types of future data. Optionally, multiple historical target data X = {y1, y2, ..., y...} are used. T As input, a shared-bottom-based multi-task learning model is used to learn the nonlinear mapping relationship between multiple historical features and multiple future features. The model contains multiple prediction sub-tasks, each predicting different influencing factors to explore valuable future feature data. That is, different sub-tasks are used to predict different types of feature data, with one sub-task predicting one type of feature data. Optionally, a long short-term memory network is used to form a shared layer in the model to couple features and mine shared temporal features among multiple factors. The specific calculation formula is shown below:

[0071] f t =σ(W f·[h t-1 x t ]+b f );

[0072]

[0073]

[0074] i t =σ(W i ·[h t-1 x t ]+b i );

[0075] o t =σ(W o ·[h t-1 x t ]+b o );

[0076] h t =o t ·tanh(C t );

[0077] Optionally, different subtasks are used to predict different types of feature data. Different fully connected layers are used for different subtasks to achieve differentiated selection of shared features. By adjusting the number of neurons in the fully connected layers, the model can be adapted to prediction tasks of different complexities, thereby improving its generalization ability to different subtasks and outputting feature data in a K-1 dimensional vector. This represents the K-1 future time-time feature data discovered, which are the load-related feature data for the predicted future times. The specific formula is as follows:

[0078]

[0079]

[0080] Optionally, the historical time-moment feature data embedding layer of the load forecasting model is constructed based on a long short-term memory network. That is, the feature data of historical time moments and the load sequence are taken as input, and the recurrent layer of the long short-term memory network structure with time series memory characteristics is used to model the complex correlation between historical feature data, and the hidden state h of the last time step is used. T The target data vector for a historical moment is obtained by using a low-dimensional vector representation of historical features.

[0081] Optionally, the load prediction layer of the prediction model is constructed based on a fully self-attention network (Transformer Block) and a multilayer perceptron. Optionally, the feature data of the predicted future time moments are used. and embedded historical feature data h T As input, a multi-task adaptive combination strategy is adopted to spontaneously focus on and combine key information from multiple sub-tasks through Transformer Block and multilayer perceptron, thereby extrapolating and predicting the data center load at future moments.

[0082] Optionally, in this embodiment of the invention, the Transformer Block, which has a simpler network structure and faster training and inference speed compared to deep learning models with residual and convolutional structures, is used to compute and extract key information from multiple input tasks in parallel. The multi-head attention mechanism involves calculating scaled dot-product attention on h different query, key, and value matrices, followed by a mapping function through matrix concatenation and linear transformation. Each set of matrices focuses on different features, and through multiple attention processes, accurate feature combination can be achieved. Optionally, the feature data predicted from multiple sub-tasks at future moments are... Transform it into a key-value matrix, and then use the target data from historical moments, i.e., the embedded historical features h. T The query matrix is ​​transformed into a scaled dot product attention matrix, which adaptively filters out the key information needed in future features and suppresses other useless information.

[0083]

[0084] The formula for calculating the attention head of the i-th scaling point product is shown above, where Q i K i and V i Let d represent the Query matrix, Key matrix, and Value matrix of the i-th attention head, respectively. ki Q represents i With K i The dimension is [dimension missing]. The multi-head attention calculation formula, consisting of multiple scaled dot product attention heads, is shown below. Where W0 represents the linear transformation coefficient matrix.

[0085] MultiHeadAttention(Q,K,V)=concat(head1,…,head h )·W0

[0086] The obtained multi-head attention output and input features h Tz1 is obtained by passing residual connections and layer normalization. z1 is then passed through two fully connected layers and residual connections, and then through layer normalization to obtain the output z2 of the Transformer Block structure. Here, W1, W2, b2, and b2 represent the weights and biases of the fully connected layers in the Transformer Block.

[0087]

[0088] z2=LayerNorm(z1+Relu(z1W1+b1)W2+b2)

[0089] Then, the z2 dimension is transformed using a multilayer perceptron to obtain the predicted load characteristics. Here, W3, W4, b3, and b4 represent the weights and biases of the fully connected layers in a multilayer perceptron.

[0090]

[0091] This enables the prediction of data center load based on target data from historical moments and characteristic data from predicted future moments, thereby improving the accuracy of data center load prediction.

[0092] For example, such as Figure 3 The data center load forecasting model shown has a structure consisting of three modules: a future feature exploration module, a historical feature embedding module, and a load forecasting module. The future feature exploration module corresponds to the future time-of-flight feature data exploration layer of the load forecasting model; the historical feature embedding module corresponds to the historical time-of-flight feature data embedding layer of the load forecasting model; and the load forecasting module corresponds to the load forecasting layer of the load forecasting model. Optionally, the future feature exploration module and the historical feature embedding module can simultaneously process the data center load data at historical time points and multiple types of load-related feature data to improve the efficiency of data center load forecasting.

[0093] The future feature exploration module explores valuable future feature data from historical influencing factors (feature data of historical moments). The load forecasting module uses the explored future feature data and the embedded historical feature data to perform multi-task adaptive combination and nonlinear transformation to predict the data center load at future moments.

[0094] The method described in the above embodiments constructs a future-moment feature data exploration layer for the load forecasting model through multi-task learning, a historical-moment feature data embedding layer for the forecasting model through a long short-term memory network, and a load forecasting layer for the forecasting model through a Transformer Block network and a multilayer perceptron. This enables the constructed forecasting model to explore valuable future features from the complex interactions among numerous historical influencing factors. Furthermore, the explored features help historical influencing factors effectively guide load forecasting, achieving data center load forecasting based on target data from historical moments and predicted feature data from future moments, thereby improving the accuracy of data center load forecasting.

[0095] In one embodiment, before obtaining the data center load prediction result by inputting data center load data from multiple historical times prior to the time to be predicted, along with multiple types of load-related feature data, into the trained load prediction model, the method further includes:

[0096] The feature data is preprocessed; the preprocessing includes: determining the maximum mutual information coefficients between each type of feature data and the data center load data, and deleting feature data whose maximum mutual information coefficients are less than the threshold.

[0097] Specifically, since there are many factors affecting data center load, that is, there are many feature data related to data center load data, and the feature data are distributed differently and the correlation between feature data is also different, treating all factors (feature data) as prediction sub-tasks to predict feature data at future times will lead to low generalization ability and training efficiency of load prediction model. Therefore, in this embodiment of the invention, K-1 factors (feature data) that are highly correlated with load data are selected for prediction to quickly improve the accuracy of load prediction.

[0098] Optionally, to further analyze the data and reduce the complexity that may arise from the large amount of feature information, this invention preprocesses the acquired historical target data. Optionally, the historical target data is first cleaned, and then feature selection is performed on the cleaned historical target data using the Maximum Information Coefficient (MIC). Feature data with a MIC value greater than or equal to 0.6 that is correlated with data center load data is selected as input for training the load prediction model. Simultaneously, feature data with a MIC value greater than or equal to 0.9 that is strongly correlated with the load is selected as sub-tasks for prediction, preventing the problem of poor model generalization ability and training speed due to an excessive number of sub-tasks.

[0099] The method described above improves the generalization ability and training efficiency of the load prediction model by preprocessing the acquired historical feature data, specifically by deleting feature data with a maximum mutual confidence coefficient less than a threshold, thereby reducing the complexity that may be caused by a large amount of feature information.

[0100] For example, in order to verify the effectiveness of the data center load forecasting method provided by the present invention, the EnergyPlus energy consumption simulation platform was used to model the data center scenario, simulating a data center area of ​​557 square meters, and the hourly load and related data of the data center were collected for one year based on real weather data, totaling 8,567 data samples. Each data sample contains 73 load features, and the specific feature descriptions are shown in Table 1.

[0101] Table 1

[0102]

[0103] To further analyze the data and reduce the complexity that might arise from the massive amount of feature information, data preprocessing is necessary. First, the data is cleaned, and then the cleaned data is used for feature selection using the Maximum Information Coefficient (MIC). Forty features with an MIC value greater than or equal to 0.6 that are correlated with data center load are selected as input for model training. Simultaneously, 15 features strongly correlated with load (MIC value greater than or equal to 0.9) are selected as sub-tasks for prediction to prevent the model's generalization ability and training speed from being poor due to an excessive number of sub-tasks. The final feature data predicted by the model is shown in Table 2. Second, the data is normalized to values ​​between 0 and 1. Then, the sequence of normalized data is transformed into a prediction of the next time point based on the previous 24 time points. Finally, the preprocessed dataset is divided into a training set (80%), a validation set (10%), and a test set (10%). In this embodiment, the number of samples N is 8567, the model input time series length T is 24, the model input dimension M is 41, and the model output dimension K is 16.

[0104] Table 2

[0105]

[0106] Next, to verify the effectiveness and universality of the proposed method, this invention selected six common load forecasting models (AR models). U VAR M LSTM U LSTM M Bi-LSTM M Transformer M) and the proposed model (FeatureExploration and Multi-task Adaptive Combination, FE-MTAC) M In comparison, this more comprehensively verifies that the proposed data center load forecasting method can effectively reduce forecasting errors.

[0107] AR U Input a one-dimensional time series autoregression (AR) model, with the input data being only historical load series.

[0108] VAR M Input a vector autoregressive (VAR) model for multivariate time series data, including historical loads and related factor sequences.

[0109] LSTM U : Input a one-dimensional time series LSTM model, with the input data being only historical load sequences.

[0110] LSTM M : Input a multivariate time series LSTM model, with input data including historical loads and related factor sequences.

[0111] Bi-LSTM M Input a multivariate time series Bi-LSTM model, with input data including historical loads and related factor sequences.

[0112] Transformer M Input a multivariate time series Transformer model, with input data including historical loads and related factor sequences.

[0113] Furthermore, this invention employs three common prediction evaluation metrics to assess model performance. MAE and RMSE can be used to quantitatively measure prediction error, while MAPE quantifies the percentage error of the predicted value relative to the true value. Their expressions are as follows:

[0114] Mean Absolute Error (MAE):

[0115]

[0116] Root Mean Square Error (RMSE):

[0117]

[0118] Mean Absolute Percentage Error (MAPE):

[0119]

[0120] Where yi represents the true value of the i-th sample. This represents the predicted value of the i-th sample.

[0121] In this embodiment, the simulation experiment was conducted on a device with 4 cores, 1.8GHz, and 16G, and the neural network was constructed using the PyTorch 1.10.2 deep learning framework. The FE-MTAC proposed in this invention was then applied. M The model was compared with six benchmark models, and the specific experimental results are shown in Table 3.

[0122] Table 3

[0123]

[0124] by AR U and VAR M LSTM U and LSTM M Experimental results from the model show that incorporating data center load-related features helps improve load forecasting accuracy, VAR. M The RMSE of the model compared to AR U Reduced by 6.74%, LSTM M The RMSE of the model compared to LSTM U It decreased by 5.48%. Meanwhile, in the aforementioned multivariate time series prediction experiments, the FE-MTAC proposed in this invention... M The model performs best compared to LSTM. M and Transformer M The RMSE decreased by 5.17% and 5.01%, respectively. This demonstrates that this embodiment, by exploring load-related future characteristics and guiding load forecasting based on historical characteristics, surpasses the traditional model's data mining capabilities and significantly improves the model's prediction accuracy.

[0125] The load forecasting device for data centers provided by the present invention will be described below. The load forecasting device for data centers described below can be referred to in correspondence with the load forecasting method for data centers described above.

[0126] Figure 4 This is a schematic diagram of the data center load forecasting device provided by the present invention. This embodiment provides a data center load forecasting device, including:

[0127] The acquisition module 710 is used to acquire target data from multiple historical moments; the target data includes data center load data and multiple types of feature data related to the load data.

[0128] The prediction module 720 is used to input the data center load of multiple historical moments before the time to be predicted and multiple types of load-related feature data into the trained load prediction model to obtain the data center load prediction result; the model is used to predict the data center load at future moments based on the target data of historical moments and the predicted feature data of future moments; the load prediction model is trained based on the target data of the first moment, the target data of the second moment, and the predicted target data of the second moment.

[0129] Optionally, the prediction module 720 is specifically used to: input the load data of the data center at historical time and multiple types of feature data related to the load data into the future time feature data exploration layer of the load prediction model to obtain the predicted feature data of the future time;

[0130] The load data of the data center at historical moments and multiple types of load data-related feature data are input into the historical moment feature data embedding layer of the load prediction model to obtain the target data vector at historical moments.

[0131] By inputting the target data vector from historical moments and the predicted feature data from future moments into the load forecasting layer of the load forecasting model, the load forecasting results for the data center are obtained.

[0132] Optionally, the prediction module 720 is specifically used to: input the target data at the first moment into the load prediction model to obtain the predicted load data of the data center at the second moment and multiple types of feature data related to the load data;

[0133] Based on the acquired target data at the second moment, the predicted data center load data at the second moment, and multiple types of load-related feature data, the load prediction model is trained to obtain the trained load prediction model.

[0134] Optionally, the prediction module 720 is specifically used to: train the load prediction model using the following loss function:

[0135]

[0136] N represents the target data samples from N data centers; K represents the dimension of the target data. This represents the predicted value of the i-th type of feature data in the j-th target data sample at time T+1. This represents the true value of the feature data of the i-th type in the j-th target data sample at time T+1.

[0137] Optionally, the future moment feature data exploration layer of the load forecasting model is constructed based on a multi-task learning network architecture and a long short-term memory network; the historical moment feature data embedding layer of the load forecasting model is constructed based on a long short-term memory network; and the load forecasting layer of the forecasting model is constructed based on a Transformer Block network and a multilayer perceptron.

[0138] The apparatus of this invention is used to execute the method in any of the foregoing method embodiments, and its implementation principle and technical effect are similar, so they will not be described again here.

[0139] Figure 5 A schematic diagram of the physical structure of an electronic device is provided. This electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840. The processor 810, communications interface 820, and memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute a data center load forecasting method. This method includes: acquiring target data from multiple historical time points; the target data includes data center load data and multiple types of load-related feature data; inputting the data center load from multiple historical time points prior to the time to be predicted and the multiple types of load-related feature data into a trained load forecasting model to obtain a data center load forecasting result; the model is used to predict the data center load at a future time point based on the target data from historical time points and the predicted feature data from future time points; the load forecasting model is trained based on the target data from a first time point, the target data from a second time point, and the predicted target data from the second time point.

[0140] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0141] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein when the program instructions are executed by a computer, the computer is able to execute the data center load forecasting method provided by the above methods, the method comprising: acquiring target data at multiple historical time points; the target data including data center load data and multiple types of load data-related feature data; inputting the data center load at multiple historical time points prior to the time to be predicted and the multiple types of load data-related feature data into a trained load forecasting model to obtain a data center load forecasting result; the model being used to forecast the data center load at a future time point based on the target data at historical time points and the predicted feature data at a future time point; the load forecasting model being trained based on the target data at a first time point, the target data at a second time point, and the predicted target data at the second time point.

[0142] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the aforementioned data center load forecasting methods. The method includes: acquiring target data from multiple historical time points; the target data including data center load data and multiple types of load data-related feature data; inputting the data center load from multiple historical time points prior to the time to be predicted and the multiple types of load data-related feature data into a trained load forecasting model to obtain a data center load forecasting result; the model being used to forecast the data center load at a future time point based on the target data from historical time points and the predicted feature data from future time points; the load forecasting model being trained based on the target data from a first time point, the target data from a second time point, and the predicted target data from the second time point.

[0143] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0144] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A load forecasting method for a data center, characterized in that, include: Acquire target data from multiple historical moments; The target data includes data center load data and multiple types of feature data related to the load data; The data center load data from multiple historical moments prior to the time to be predicted, along with various types of feature data related to the load data, are input into the trained load prediction model to obtain the data center load prediction result. The model is used to predict the data center load at future moments based on the target data from historical moments and the predicted feature data from future moments. The load prediction model is trained based on the target data at the first time, the target data at the second time, and the predicted target data at the second time. The process involves inputting data center load data from multiple historical moments prior to the predicted time, along with various types of feature data related to the load data, into a trained load prediction model to obtain data center load prediction results. This includes: The load data of the data center at historical moments and multiple types of feature data related to the load data are input into the future moment feature data exploration layer of the load prediction model to obtain the predicted feature data for the future moment. The load data of the data center at historical time and multiple types of feature data related to the load data are input into the historical time feature data embedding layer of the load prediction model to obtain the target data vector at historical time. The target data vector of the historical moment and the feature data of the predicted future moment are input into the load prediction layer of the load prediction model to obtain the load prediction result of the data center. Before inputting the data center load data from multiple historical moments prior to the predicted time and multiple types of feature data related to the load data into the trained load prediction model to obtain the data center load prediction result, the method further includes: The target data at the first moment is input into the load prediction model to obtain the predicted data center load data at the second moment and multiple types of feature data related to the load data. Based on the acquired target data at the second moment, the predicted data center load data at the second moment, and multiple types of feature data related to the load data, the load prediction model is trained to obtain the trained load prediction model. The process of training the load prediction model based on the acquired target data at the second time moment, the predicted data center load data at the second time moment, and multiple types of feature data related to the load data, to obtain the trained load prediction model, includes: The load forecasting model is trained using the following loss function: ； N represents the target data samples from N data centers; K represents the dimension of the target data. express The predicted value of the feature data of the i-th type in the j-th target data sample at time j. express The true value of the feature data of the i-th type in the j-th target data sample at time j; The future moment feature data exploration layer of the load prediction model is constructed based on a multi-task learning network architecture and a long short-term memory network; the historical moment feature data embedding layer of the load prediction model is constructed based on a long short-term memory network; and the load prediction layer of the prediction model is constructed based on a fully self-attention network and a multilayer perceptron. Before inputting the data center load data from multiple historical moments prior to the predicted time and multiple types of feature data related to the load data into the trained load prediction model to obtain the data center load prediction result, the method further includes: The feature data is preprocessed; the preprocessing includes: determining the maximum mutual information coefficients between each type of feature data and the load data of the data center, and deleting feature data whose maximum mutual information coefficients are less than a threshold.

2. A load forecasting apparatus for a data center used to implement the load forecasting method of claim 1, characterized in that, include: The acquisition module is used to acquire target data from multiple historical moments. The target data includes data center load data and multiple types of feature data related to the load data; The prediction module is used to input the data center load of multiple historical moments before the time to be predicted and multiple types of feature data related to the load data into the trained load prediction model to obtain the data center load prediction result; the model is used to predict the data center load at future moments based on the target data at historical moments and the predicted feature data at future moments. The load prediction model is trained based on the target data at the first time step, the target data at the second time step, and the predicted target data at the second time step.

3. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the data center load forecasting method as described in claim 1.

4. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the data center load forecasting method as described in claim 1.

5. A computer program product having executable instructions stored thereon, characterized in that, When executed by the processor, this instruction causes the processor to implement the data center load forecasting method as described in claim 1.

Images