Data center energy consumption prediction method and system based on large language model

Abstract

The invention provides a data center energy consumption prediction method and system based on a large language model, and relates to the technical field of data center energy consumption prediction. According to the method, based on SHAP analysis, input features are quantized and sorted, features having key influences on energy consumption prediction are screened and measured, and specific contributions of the input features to prediction results are quantized. Meanwhile, the improved self-adaptive SHAP analysis method can realize dynamic updating of feature importance so as to adapt to the change of the operating environment and load of the data center. And according to the automatically updated key features, constructing a knowledge graph, and revealing causal and collaborative relationships among elements such as equipment, environment, load and the like. And then, automatically generating a structured prompt text by using the key features, an SHAP analysis result and professional domain knowledge in the knowledge graph, and dynamically guiding a large language model to pay attention to and deduce the most influential features and the interaction relationship thereof, thereby realizing more accurate data center energy consumption prediction.

Description

Technical Field

[0001] This invention relates to the field of data center energy consumption prediction technology, specifically to a data center energy consumption prediction method and system based on a large language model. Background Technology

[0002] With the continuous expansion of data center scale and the increasing demand for energy conservation and emission reduction, how to accurately and effectively predict their energy consumption has become an urgent technical problem to be solved.

[0003] Currently, existing technologies mainly employ traditional regression models or deep learning models to predict data center energy consumption. For example, a short-term cooling energy consumption prediction model based on SEN-MH-BiLSTM can better uncover the dependency between data center cooling energy consumption and feature sequences. Alternatively, a data center energy consumption prediction method based on DCIM can generate a dynamic topology matrix by simultaneously acquiring parameters from multiple modules and aligning them spatiotemporally. This matrix is ​​then combined with energy propagation weights and historical event signals to generate a prediction map, and parameters are adjusted based on the results to output cooling and power supply regulation commands.

[0004] However, the aforementioned existing technologies all struggle to quantify the specific contribution of input features to the prediction results, leading to poor model interpretability. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a data center energy consumption prediction method and system based on a large language model, solving the technical problem that existing technologies struggle to quantify the specific contribution of input features to prediction results.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a data center energy consumption prediction method based on a large language model, comprising: Acquire multi-source heterogeneous characteristic data closely related to energy consumption in the data center, and preprocess the multi-source heterogeneous characteristic data; An improved adaptive SHAP analysis method is used to calculate the SHAP value of each feature in the preprocessed multi-source heterogeneous feature data. The average absolute SHAP value of each feature is calculated based on the SHAP value of each feature. The features are ranked and key features are selected using the average absolute SHAP value as the importance criterion. Construct a knowledge graph of data center energy consumption based on key features; Construct prompt text based on key features, SHAP analysis results, and professional knowledge from the knowledge graph; The prompt text is processed by a pre-trained large language model, and the energy consumption prediction results are output in text form. The execution flow of the improved adaptive SHAP analysis method is as follows: Real-time monitoring of robustness scores; if a robustness score exceeds a preset threshold, the SHAP values ​​of each feature are immediately recalculated. Furthermore, a fixed time interval is set; each time the interval is reached, the current data within that interval is automatically extracted, and the SHAP values ​​of each feature in the current data are recalculated. The robustness score is calculated using the following methods: in, It is a feature i A steady scoring performance It is a feature i The average absolute SHAP value, It is a feature i The variance of the SHAP value, It is the penalty coefficient.

[0007] Preferably, the multi-source heterogeneous feature data includes environmental parameters, temperature sensor data, fan operating status, power module indicators, and PCIe bus related information.

[0008] Preferably, the SHAP value of each feature is calculated in the following ways: The SHAP value for each feature is calculated using TreeExplainer from the SHAP method package. The calculation formula is as follows: in, Represents the set of all features. For features not included Feature subset, and They are respectively and The number of features, For the model in feature subset The prediction results above ) is a feature that includes The model prediction results after that.

[0009] Preferably, the step of constructing prompt text based on key features, SHAP analysis results, and professional knowledge in the knowledge graph includes: Role-setting for large language models; Provide detailed explanations of key features; The key features and their impact directions obtained from SHAP analysis are embedded into the prompt text in a structured manner, so that the prompt text clearly points out the trend of the impact of changes in each feature on energy consumption; and the professional knowledge and causal relationships in the knowledge graph are incorporated into the prompt text as reasoning prompts.

[0010] Preferably, the step of constructing prompt text based on key features, SHAP analysis results, and professional knowledge in the knowledge graph further includes: Whenever a change in key features or a change in the ranking of key features is detected, the prompt text content is automatically reconstructed and updated.

[0011] Preferably, the prompt text also includes prompt information that guides the large language model to focus on higher-order interaction relationships between features.

[0012] Preferably, the pre-trained large language model is fine-tuned using LoRA parameter efficient fine-tuning technology during the training process.

[0013] Secondly, this invention provides a data center energy consumption prediction system based on a large language model, comprising: The data acquisition and preprocessing module is used to acquire multi-source heterogeneous characteristic data closely related to energy consumption in the data center, and to preprocess the multi-source heterogeneous characteristic data. The feature evaluation and screening module uses an improved adaptive SHAP analysis method to calculate the SHAP value of each feature in the preprocessed multi-source heterogeneous feature data, and calculates the average absolute SHAP value of each feature based on the SHAP value of each feature; using the average absolute SHAP value as the importance criterion, the features are sorted and key features are screened out. The knowledge graph construction module is used to build a knowledge graph of data center energy consumption based on key features; The prompt text construction module is used to build prompt text based on key features, SHAP analysis results, and professional knowledge in the knowledge graph; The energy consumption prediction and result output module is used to process the prompt text through a pre-trained large language model and output the energy consumption prediction results in text form. The execution flow of the improved adaptive SHAP analysis method is as follows: Real-time monitoring of robustness scores; if a robustness score exceeds a preset threshold, the SHAP values ​​of each feature are immediately recalculated. Furthermore, a fixed time interval is set; each time the interval is reached, the current data within that interval is automatically extracted, and the SHAP values ​​of each feature in the current data are recalculated. The robustness score is calculated using the following methods: in, It is a feature iA steady scoring performance It is a feature i The average absolute SHAP value, It is a feature i The variance of the SHAP value, It is the penalty coefficient.

[0014] Thirdly, the present invention provides a computer-readable storage medium storing a computer program for predicting data center energy consumption based on a large language model, wherein the computer program causes a computer to execute the data center energy consumption prediction method based on a large language model as described above.

[0015] Fourthly, the present invention provides an electronic device, comprising: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing data center energy consumption prediction based on large language models as described above.

[0016] (III) Beneficial Effects This invention provides a method and system for predicting data center energy consumption based on a large language model. Compared with existing technologies, it has the following advantages: This invention, based on SHAP analysis, quantifies and ranks input features, filters and measures features with a key impact on energy consumption prediction, and quantifies the specific contribution of input features to the prediction results. Simultaneously, the improved adaptive SHAP analysis method enables dynamic updates of feature importance to adapt to changes in the data center operating environment and load. Based on the automatically updated key features, these are embedded as prior knowledge into the model inference process, enhancing the interpretability of the model's impact on key features, thereby achieving more accurate data center energy consumption prediction. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a block diagram of a data center energy consumption prediction method based on a large language model, according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This application provides a data center energy consumption prediction method and system based on a large language model, which solves the technical problem that it is difficult to quantify the specific contribution of input features to the prediction results in the prior art, and realizes the interpretability of the model's influence on key features.

[0021] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows: The existing methods have the following main drawbacks: 1. The model has poor interpretability and it is difficult to quantify the specific contribution of input features to the prediction results. Most existing energy consumption prediction methods cannot automatically select, prioritize, and dynamically inject important features based on the actual contribution of features to the prediction results. This makes it difficult to balance prediction accuracy and interpretability when the model is subjected to high-dimensional redundancy or even noisy features; moreover, manual selection and adjustment of features is not only time-consuming and laborious, but also prone to missing the dynamic changes of key features.

[0022] 2. Insufficient mining of correlations among multi-source heterogeneous high-dimensional features and limited utilization of contextual information. Existing technologies mostly use traditional regression methods or deep learning models to predict data center energy consumption. When faced with multi-source heterogeneous, high-dimensional, and complexly correlated input features in real-world scenarios, the model structure is limited, making it difficult to continuously perceive and self-adjust the importance of features and their contextual impact. This results in difficulty in actively mining and utilizing the latest effective information.

[0023] 3. Large parameter scale, high training cost and limited generalization ability. Although deep learning methods have improved the ability to process high-dimensional and complex features, they usually involve a large number of parameters and complex network structures, are highly dependent on large-scale high-quality data, have high computational costs for training and inference, and are highly dependent on manual hyperparameter tuning.

[0024] To address the aforementioned issues, this invention proposes a data center energy consumption prediction method and system based on a large language model. Addressing problems such as insufficient utilization of multi-source heterogeneous high-dimensional feature associations, low model parameter efficiency, and poor interpretability in data center energy consumption prediction, this method proposes a large language model-based energy consumption prediction approach that combines context enhancement and efficient parameter fine-tuning. First, this method uses the SHAP method and drift detection to dynamically measure and adaptively filter the contribution of features, identifying key features affecting energy consumption in the current scenario. Based on these key features, a knowledge graph is constructed to reveal the causal and synergistic relationships between elements such as equipment, environment, and load. Subsequently, relying on feature definition, SHAP analysis, and professional knowledge from the knowledge graph, structured prompt text is automatically generated to dynamically guide the large language model to focus on and reason about the most influential features and their interactions. Finally, by utilizing the large language model and LORA parameter efficient fine-tuning technology, the method can fully utilize the contextual relationships between multi-source, high-dimensional features and improve model interpretability and parameter utilization efficiency, thereby achieving more accurate data center energy consumption prediction.

[0025] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0026] This invention provides a data center energy consumption prediction method based on a large language model, such as... Figure 1 As shown, it includes: S1. Obtain multi-source heterogeneous characteristic data closely related to energy consumption in the data center, and preprocess the multi-source heterogeneous characteristic data; S2. The improved adaptive SHAP analysis method is used to calculate the SHAP value of each feature in the preprocessed multi-source heterogeneous feature data. The average absolute SHAP value of each feature is calculated based on the SHAP value of each feature. The features are ranked and key features are selected using the average absolute SHAP value as the importance criterion. S3. Construct a knowledge graph of data center energy consumption based on key features; S4. Construct prompt text based on key features, SHAP analysis results, and professional knowledge in the knowledge graph; S5. Process the prompt text using a pre-trained large language model and output the energy consumption prediction results in text form. The execution flow of the improved adaptive SHAP analysis method is as follows: Real-time monitoring of robustness scores; if a robustness score exceeds a preset threshold, the SHAP values ​​of each feature are immediately recalculated. Furthermore, a fixed time interval is set; each time the interval is reached, the current data within that interval is automatically extracted, and the SHAP values ​​of each feature in the current data are recalculated. The robustness score is calculated using the following methods: in, It is a feature i A steady scoring performance It is a feature i The average absolute SHAP value, It is a feature i The variance of the SHAP value, It is the penalty coefficient.

[0027] This invention, based on SHAP analysis, quantifies and ranks input features, filters and measures features with a key impact on energy consumption prediction, and quantifies the specific contribution of input features to the prediction results. Simultaneously, the improved adaptive SHAP analysis method enables dynamic updates of feature importance to adapt to changes in the data center operating environment and load. Based on the automatically updated key features, these are embedded as prior knowledge into the model inference process, enhancing the interpretability of the model's impact on key features, thereby achieving more accurate data center energy consumption prediction.

[0028] The following is a detailed description of each step: In one embodiment, S1, acquire multi-source heterogeneous characteristic data closely related to energy consumption in the data center, and preprocess the multi-source heterogeneous characteristic data. The specific implementation process is as follows: First, multi-source heterogeneous data closely related to energy consumption in the data center are collected, covering multiple dimensions such as environmental parameters, temperature sensor data, fan operating status, power module indicators, and PCIe bus related information.

[0029] Data acquisition is performed in 15-minute intervals to ensure continuous and detailed dynamic monitoring of the data center's operational status. During the data preprocessing stage, normalization methods are used to standardize the raw data to address the differences in the dimensions and distributions of various features. This eliminates scale effects between different features, improving the effectiveness of subsequent modeling and the convergence speed of model training.

[0030] In one embodiment, S2, an improved adaptive SHAP analysis method is used to calculate the SHAP value of each feature in the preprocessed multi-source heterogeneous feature data, and the average absolute SHAP value of each feature is calculated based on the SHAP value of each feature; the average absolute SHAP value is used as the importance criterion to rank the features and select key features. The specific implementation process is as follows: For preprocessed multi-source heterogeneous feature data, the SHAP analysis method is used to evaluate the contribution of each input feature to the energy consumption prediction task. Furthermore, by incorporating seasonal changes and variations in operating conditions such as load levels, automatic feature importance assessment triggered by data distribution drift is achieved. First, all preprocessed features are input into the XGBoost model, and model training is completed on the training set. Based on the trained model, the SHAP value of each feature is calculated using the TreeExplainer method from the SHAP package. This is to quantify its marginal contribution to the model's prediction results.

[0031] in, Represents the set of all features. For features not included Feature subset, and They are respectively and The number of features, For the model in feature subset The prediction results above ) is a feature that includes The model prediction results after that.

[0032] By traversing all possible feature subsets, the calculation is performed on features that are included or not included. The model output changes under different conditions, and features are obtained by weighted averaging. The SHAP value of each feature is calculated, and the average absolute SHAP value of each feature is further calculated as a measure of global feature importance.

[0033] in, It is a feature i The average absolute SHAP value, For the sample size, For the first The first sample The SHAP value of each feature.

[0034] In this embodiment of the invention, the SHAP value is dynamically changing. The recalculation of the SHAP value for each feature is achieved through the following two triggering mechanisms, ensuring the timeliness and reliability of feature importance assessment: 1. Threshold triggering mechanism: When the robustness score exceeds a preset threshold (in this embodiment, the threshold is set to the 95th percentile), the recalculation process of the SHAP value for each feature is immediately initiated, ensuring rapid updates to the assessment results when there are significant fluctuations in data distribution or feature influence; 2. Timed triggering mechanism: Every hour (i.e., a fixed interval of 1 hour), the SHAP value of each feature is recalculated for the data within the current hour, thereby dynamically reflecting the trend of feature importance changes "at this moment." Specifically: Given the highly dynamic nature of the data center operating environment, the impact of certain characteristics on energy consumption varies with changes in the load temperature of the CPU, graphics card, and server, as well as the ambient temperature. For example, during hot seasons, the impact of ambient temperature on energy consumption is more significant; and when hardware heat dissipation efficiency decreases, the importance of certain hardware indicators also increases. Simultaneously, the mean and variance of environmental variables are calculated and combined with a preset penalty coefficient to form a comprehensive robustness score.

[0035] in, It is a feature i A steady scoring performance It is a feature i The average absolute SHAP value, It is a feature i The variance of the SHAP value, This is the penalty coefficient, which is usually set to 0.3.

[0036] If the robust score exceeds the preset 95th percentile as a threshold, the system automatically determines that the data distribution has drifted and immediately triggers a new round of feature importance assessment process.

[0037] Based on the newly calculated importance scores of each feature, the top K key features with a cumulative importance of 90% are selected as the core variables for subsequent analysis and modeling.

[0038] In one embodiment, S3, a knowledge graph of data center energy consumption is constructed based on key features. The specific implementation process is as follows: Based on these key characteristics, a knowledge graph covering data center energy consumption equipment, environment, load, and operational behavior is further constructed, including the following elements: Physical nodes: ambient temperature, CPU load, fan speed, power consumption, etc. Attribute nodes: numerical values, time, category, etc. of features; Relationship edges: influence, attribution, regulation, accompaniment, etc., clarifying the causal, correlation, and synergistic relationships between various features; Rules and Facts: High ambient temperature → high energy consumption; high CPU utilization + high fan speed → high energy consumption.

[0039] Defining these relationships provides structured support for subsequent knowledge enhancement.

[0040] In one embodiment, step S4 involves constructing prompt text based on key features, a knowledge graph, and the reasoning process. The specific implementation process is as follows: After obtaining key features and their knowledge graph, a structured prompt text automatic generation method was designed to improve the accuracy of LLM in data center energy consumption prediction tasks. First, the model is given a role: a data center energy consumption prediction expert. Second, key features are explained in detail, including their physical meaning, units, typical value ranges, and other basic information. Then, the key features obtained from SHAP analysis and their impact directions are embedded into the prompt text in a structured manner. This ensures that the prompt text clearly indicates the trend of each feature's change on energy consumption, guiding LLM to focus on these key variables during inference. Simultaneously, professional knowledge and causal relationships from the knowledge graph are incorporated into the inference prompts, such as "high ambient temperature usually leads to high energy consumption" and "increased CPU utilization may increase cooling pressure," enhancing LLM's contextual understanding and professional inference capabilities. Furthermore, the prompt text guides LLM to focus on higher-order interactions between features, analyzing the comprehensive impact of changes in factors such as input power and load on energy consumption, thereby improving the scientific rigor and accuracy of the inference. Finally, whenever a change in key features or a change in the ranking of key features is detected, the Prompt content is automatically reconstructed and updated to ensure that LLM always performs energy consumption prediction and analysis based on the latest feature importance and knowledge graph.

[0041] In one embodiment, S5, the prompt text is processed using a pre-trained large language model, and the energy consumption prediction result is output in text form. The specific implementation process is as follows: It should be noted that while pre-trained large language models possess rich language understanding and expression capabilities on large-scale general corpora, they often fail to achieve ideal results when directly applied to specific tasks such as data center energy consumption prediction. Therefore, supervised fine-tuning (SFT) based on structured prompt text is used to improve the model's adaptability and prediction accuracy within the domain. To further improve fine-tuning efficiency and reduce computational resource consumption, a LoRA parameter efficient fine-tuning technique is introduced into the training process of the large language model. This method inserts two trainable low-rank matrices A and B into the original model structure, projects the input data to a low-dimensional space, and then restores it to the original dimension. The transformed result is then added to the original output, achieving efficient model adaptation without directly adjusting a large-scale weight matrix, requiring only a small number of parameter updates. During the fine-tuning process, the model not only learns the complex relationship between multi-dimensional features and energy consumption but also fully absorbs the importance of each feature and its contextual information, thereby better identifying key variables and improving prediction accuracy and interpretability.

[0042] The pre-trained large language model processes the prompt text and outputs the energy consumption prediction results in text form.

[0043] This invention provides a data center energy consumption prediction system based on a large language model, comprising: The data acquisition and preprocessing module is used to acquire multi-source heterogeneous characteristic data closely related to energy consumption in the data center, and to preprocess the multi-source heterogeneous characteristic data. The feature evaluation and screening module uses an improved adaptive SHAP analysis method to calculate the SHAP value of each feature in the preprocessed multi-source heterogeneous feature data, and calculates the average absolute SHAP value of each feature based on the SHAP value of each feature; using the average absolute SHAP value as the importance criterion, the features are sorted and key features are screened out. The knowledge graph construction module is used to build a knowledge graph of data center energy consumption based on key features; The prompt text construction module is used to build prompt text based on key features, SHAP analysis results, and professional knowledge in the knowledge graph; The energy consumption prediction and result output module is used to process the prompt text through a pre-trained large language model and output the energy consumption prediction results in text form. The execution flow of the improved adaptive SHAP analysis method is as follows: Real-time monitoring of robustness scores; if a robustness score exceeds a preset threshold, the SHAP values ​​of each feature are immediately recalculated. Furthermore, a fixed time interval is set; each time the interval is reached, the current data within that interval is automatically extracted, and the SHAP values ​​of each feature in the current data are recalculated. The robustness score is calculated using the following methods: in, It is a feature i A steady scoring performance It is a feature i The average absolute SHAP value, It is a feature i The variance of the SHAP value, It is the penalty coefficient.

[0044] It is understood that the data center energy consumption prediction system based on large language models provided in this embodiment of the invention corresponds to the data center energy consumption prediction method based on large language models described above. The explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding content in the data center energy consumption prediction method based on large language models, and will not be repeated here.

[0045] This invention also provides a computer-readable storage medium storing a computer program for predicting data center energy consumption based on a large language model, wherein the computer program causes a computer to execute the data center energy consumption prediction method based on a large language model as described above.

[0046] This invention also provides an electronic device, including: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the data center energy consumption prediction method based on a large language model as described above.

[0047] In summary, compared with existing technologies, it has the following beneficial effects: 1. This invention, based on SHAP analysis, quantifies and ranks input features, filters and measures features with key impact on energy consumption prediction, and quantifies the specific contribution of input features to the prediction results. Simultaneously, the improved adaptive SHAP analysis method enables dynamic updating of feature importance to adapt to changes in the data center operating environment and load. Based on the automatically updated key features, these are embedded as prior knowledge into the model inference process, thereby enhancing the interpretability of the model regarding the impact of key features.

[0048] 2. By embedding feature definitions, SHAP analysis, and professional knowledge from knowledge graphs into prompt text in a structured manner, and by providing reasoning suggestions and decision guidance, large models can focus on key variables, clarify causal relationships, and improve the scientific nature of the model's reasoning and the accuracy of its predictions.

[0049] 3. The adaptive SHAP analysis method of this application can automatically and periodically evaluate the importance of features and update them in real time, so that the large language model can continuously pay attention to the latest information and feature changes and actively adapt to the dynamic changes of the environment and business scenarios.

[0050] 4. Employing LoRA parameter fine-tuning technology, only a small number of parameters need to be fine-tuned, without the need for large-scale adjustments to the original weight matrix, thus reducing the computational resource consumption during model training and inference.

[0051] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0052] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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. Such 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 data center energy consumption prediction method based on a large language model, characterized in that, include: Acquire multi-source heterogeneous characteristic data closely related to energy consumption in the data center, and preprocess the multi-source heterogeneous characteristic data; An improved adaptive SHAP analysis method is used to calculate the SHAP value of each feature in the preprocessed multi-source heterogeneous feature data. The average absolute SHAP value of each feature is calculated based on the SHAP value of each feature. The features are ranked and key features are selected using the average absolute SHAP value as the importance criterion. Construct a knowledge graph of data center energy consumption based on key features; Construct prompt text based on key features, SHAP analysis results, and professional knowledge from the knowledge graph; The prompt text is processed by a pre-trained large language model, and the energy consumption prediction results are output in text form. The execution flow of the improved adaptive SHAP analysis method is as follows: Real-time monitoring of robustness scores; if a robustness score exceeds a preset threshold, the SHAP values ​​of each feature are immediately recalculated. Furthermore, a fixed time interval is set; each time the interval is reached, the current data within that interval is automatically extracted, and the SHAP values ​​of each feature in the current data are recalculated. The robustness score is calculated using the following methods: in, It is a feature i A steady scoring performance It is a feature i The average absolute SHAP value, It is a feature i The variance of the SHAP value, It is the penalty coefficient.

2. The data center energy consumption prediction method based on a large language model as described in claim 1, characterized in that, The multi-source heterogeneous feature data includes environmental parameters, temperature sensor data, fan operating status, power module indicators, and PCIe bus related information.

3. The data center energy consumption prediction method based on a large language model as described in claim 1, characterized in that, The SHAP values ​​of each feature are calculated using the following methods: The SHAP value for each feature is calculated using TreeExplainer from the SHAP method package. The calculation formula is as follows: in, Represents the set of all features. For features not included Feature subset, and They are respectively and The number of features, For the model in feature subset The prediction results above ) is a feature that includes The model prediction results after that.

4. The data center energy consumption prediction method based on a large language model as described in claim 1, characterized in that, The construction of prompt text based on key features, SHAP analysis results, and professional knowledge in the knowledge graph includes: Role-setting for large language models; Provide detailed explanations of key features; The key features and their impact directions obtained from SHAP analysis are embedded into the prompt text in a structured manner, so that the prompt text clearly points out the trend of the impact of changes in each feature on energy consumption; and the professional knowledge and causal relationships in the knowledge graph are incorporated into the prompt text as reasoning prompts.

5. The data center energy consumption prediction method based on a large language model as described in claim 4, characterized in that, The method of constructing prompt text based on key features, SHAP analysis results, and professional knowledge in the knowledge graph also includes: Whenever a change in key features or a change in the ranking of key features is detected, the prompt text content is automatically reconstructed and updated.

6. The data center energy consumption prediction method based on a large language model as described in claim 4, characterized in that, The prompt text also includes prompts that guide the large language model to focus on higher-order interaction relationships between features.

7. The data center energy consumption prediction method based on a large language model as described in any one of claims 1 to 6, characterized in that, The pre-trained large language model is fine-tuned using LoRA parameter efficient fine-tuning technology during the training process.

8. A data center energy consumption prediction system based on a large language model, characterized in that, include: The data acquisition and preprocessing module is used to acquire multi-source heterogeneous characteristic data closely related to energy consumption in the data center, and to preprocess the multi-source heterogeneous characteristic data. The feature evaluation and screening module uses an improved adaptive SHAP analysis method to calculate the SHAP value of each feature in the preprocessed multi-source heterogeneous feature data, and calculates the average absolute SHAP value of each feature based on the SHAP value of each feature; using the average absolute SHAP value as the importance criterion, the features are sorted and key features are screened out. The knowledge graph construction module is used to build a knowledge graph of data center energy consumption based on key features; The prompt text construction module is used to build prompt text based on key features, SHAP analysis results, and professional knowledge in the knowledge graph; The energy consumption prediction and result output module is used to process the prompt text through a pre-trained large language model and output the energy consumption prediction results in text form. The execution flow of the improved adaptive SHAP analysis method is as follows: Real-time monitoring of robustness scores; if a robustness score exceeds a preset threshold, the SHAP values ​​of each feature are immediately recalculated. Furthermore, a fixed time interval is set; each time the interval is reached, the current data within that interval is automatically extracted, and the SHAP values ​​of each feature in the current data are recalculated. The robustness score is calculated using the following methods: in, It is a feature i A steady scoring performance It is a feature i The average absolute SHAP value, It is a feature i The variance of the SHAP value, It is the penalty coefficient.

9. A computer-readable storage medium, characterized in that, It stores a computer program for predicting data center energy consumption based on a large language model, wherein the computer program causes a computer to execute the data center energy consumption prediction method based on a large language model as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing data center energy consumption prediction based on large language models as described in any one of claims 1 to 7.

Images