A cooling system energy consumption prediction method based on E-FuzzyKAN

Abstract

This invention discloses a cooling system energy consumption prediction method based on E-FuzzyKAN, comprising the following steps: collecting multi-source data from the data center cooling system, preprocessing it to construct a model data sample set; constructing an E-FuzzyKAN energy consumption prediction model composed of a fuzzy system and a KAN; dividing the data sample set into a training set and a test set, training the constructed E-FuzzyKAN model using the training set, and verifying the model's predictive performance using the test set; inputting real-time collected cooling system operating data into the trained model, and outputting accurate energy consumption prediction results. This invention significantly improves the accuracy of data center cooling system energy consumption prediction by integrating the synergistic advantages of fuzzy systems and KAN networks, providing data support for the dynamic optimization and adjustment of cooling system setpoints, ensuring stable operation of the data center at a suitable temperature, and effectively avoiding energy waste caused by excessive cooling.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and specifically to a method for predicting the energy consumption of a cooling system based on E-FuzzyKAN. Background Technology

[0002] With the rapid expansion of digital infrastructure, the energy consumption of data centers continues to rise. Cooling systems, as a core supporting facility ensuring stable server operation, account for approximately 30% of the total energy consumption of a data center, making them a key control target for achieving green and energy-efficient data centers. Practice shows that cooling system energy consumption is not independent; it is strongly coupled with the energy consumption fluctuations of the server cluster and is directly affected by two core control parameters: supply air temperature and supply air volume. Dynamic changes in server energy consumption alter the heat load distribution within the data center, and the regulation of supply air temperature and volume is the core means for the cooling system to match the heat load. Therefore, accurately capturing the correlation among these three factors is a crucial prerequisite for achieving high-precision prediction of cooling system energy consumption.

[0003] Traditional physical models for predicting cooling system energy consumption are mostly based on the assumption of a fixed heat load. However, in actual data centers, the utilization rate and business load of server clusters fluctuate dynamically over time, resulting in significant time-varying characteristics in server energy consumption, and the heat load within the data center changes in real time accordingly. Meanwhile, supply air temperature and volume, as control parameters of the cooling system, are closely related to its energy consumption. While conventional machine learning prediction models, such as Support Vector Machines and Long Short-Term Memory Networks, can mine patterns in cooling system energy consumption changes through time-series data, they have significant shortcomings in feature processing. Most models do not use server energy consumption fluctuations as core input features, relying solely on the cooling system's own operating parameters for prediction, ignoring the uncertainty of heat load changes at the energy source.

[0004] Existing cooling system energy consumption prediction technologies have significant technical bottlenecks in the collaborative analysis of server energy consumption fluctuations and core air supply control parameters, making it difficult to achieve high-precision energy consumption prediction under complex operating conditions. Summary of the Invention

[0005] 1. The technical problem to be solved:

[0006] How to accurately predict the relationship between server energy consumption fluctuations and air supply parameters, so as to overcome the limitations of existing technology and provide a reliable quantitative basis for intelligent control and energy efficiency optimization of data center cooling systems.

[0007] 2. Technical Solution:

[0008] To address the above problems, this invention provides a cooling system energy consumption prediction method based on E-FuzzyKAN, characterized by the following steps:

[0009] Step 1: Collect operational data that affects the energy consumption of the data center cooling system, including server energy consumption, air supply temperature, and air supply volume; preprocess the operational data to obtain a model data sample set.

[0010] Step 2: Construct a cooling system energy consumption prediction model based on E-FuzzyKAN.

[0011] Step 3: Divide the data sample set into a training set and a test set. Use the training set to train the E-FuzzyKAN prediction model built in Step 2, and use the test set to verify the prediction performance of the model, thus obtaining the trained E-FuzzyKAN-based cooling system energy consumption prediction model.

[0012] Step 4: Input the real-time collected operating data into the trained E-FuzzyKAN prediction model for prediction, and output the final energy consumption prediction result of the cooling system.

[0013] Furthermore, in step 1, different sensors are deployed in the computer room, and based on the energy transfer logic of server heat dissipation and cooling system refrigeration, operational data that has a significant impact on the energy consumption of the cooling system are selected.

[0014] Furthermore, the preprocessing of the running data includes: removal of historical time series outliers, filling in missing values, and data normalization.

[0015] Furthermore, in step 2: by organically integrating the structured logical expression capability of the fuzzy reasoning system with the efficient nonlinear learning characteristics of the KAN (Kolmogorov-Arnold Network) neural network, a synergistically enhanced prediction framework is formed.

[0016] Furthermore, the specific method for forming a collaboratively enhanced prediction framework is as follows:

[0017] Step 21: Use the input sequence X as the input variable of the network, where each node represents an input variable. The specific formula is as follows:

[0018] (1)

[0019] in, The first The input number The mean and standard deviation of a fuzzy set. For the number of input variables, The number of fuzzy sets, For parameterized activation functions, This represents the degree of membership.

[0020] Step 22: The fuzzy rule layer implements fuzzy inference through fuzzy rules. The output is the activation strength of the fuzzy rule, which is obtained by multiplying the membership degrees of each premise part. The specific formula is as follows:

[0021] (2)

[0022] in, For the first The activation strength of a fuzzy rule It is a fuzzy rule number.

[0023] Step 23: The normalization layer normalizes the output of the fuzzy rule layer so that the sum of the activation intensities of all fuzzy rules is one. The specific formula is as follows:

[0024] (3)

[0025] in, This is the normalized matrix.

[0026] Step 24: After normalizing the activation intensity, reshape it into a structured tensor representing the fuzzy rule response of each feature, as shown in the following formula:

[0027] (4)

[0028] Among them, among them, For fuzzy rule enhancement matrix, For two one-dimensional convolution operations, This is a pooling operation.

[0029] Step 25: The output layer calculates the weighted sum of the outputs of the enhancement layer, outputting the predicted cooling system energy consumption result. The specific formula is as follows:

[0030] (5)

[0031] in, For linear weights, For bias, This is a predicted value for the energy consumption of the cooling system.

[0032] 3. Beneficial effects:

[0033] This invention effectively addresses the shortcomings of traditional prediction techniques, such as insufficient accuracy and poor robustness under complex operating conditions, and offers several significant advantages. It not only rapidly approximates complex nonlinear systems but also effectively handles uncertain input data through fuzzy set theory, fundamentally reducing data uncertainty interference. Furthermore, it enhances the semantic expression and feature extraction capabilities of rules. Results show that, based on the high-precision energy consumption prediction results, the proposed energy consumption prediction model allows data center operators to obtain the changing trends of cooling system energy consumption under different server load scenarios, providing accurate and reliable data support for the dynamic optimization and adjustment of core setpoints such as cooling system air supply temperature and air volume. Attached Figure Description

[0034] Figure 1 This is a flowchart of the method.

[0035] Figure 2 This is a diagram illustrating the operational logic of the computer room involved in this method.

[0036] Figure 3 This is a network diagram for predicting the energy consumption of the cooling system involved in this method.

[0037] Figure 4 The comparison chart shows the predicted energy consumption of the data center cooling system in the example to verify the present invention. Detailed Implementation

[0038] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0039] A method for predicting the energy consumption of a cooling system based on E-FuzzyKAN, characterized by the following steps:

[0040] Step 1: Obtain operational data affecting the energy consumption of the data center cooling system based on the data center's operating logic and layout. The operational data includes server energy consumption, air supply temperature, and air supply volume. Preprocess the operational data to obtain a model data sample set.

[0041] Step 2: Construct a cooling system energy consumption prediction model based on E-FuzzyKAN. The E-FuzzyKAN method organically integrates the structured logical expression capability of the fuzzy inference system with the efficient nonlinear learning characteristics of the KAN neural network to form a synergistically enhanced prediction framework. On the one hand, this method leverages the powerful nonlinear fitting capability of the KAN network to quickly approximate the complex dynamic characteristics related to the energy consumption of the cooling system. On the other hand, it uses fuzzy set theory to model and process the uncertainties in the input data, such as sensor noise and temperature fluctuations, reducing the impact of data disturbances on prediction accuracy. The designed enhancement layer combines the sliding convolution operation of the convolution kernel on the input data, and uses shared convolution kernels to learn the local interactions between fuzzy rules, which can capture useful common activations and suppress redundant or conflicting rules.

[0042] Step 3: Divide the data sample set into a training set and a test set. Use the training set to train the E-FuzzyKAN prediction model built in Step 2, and use the test set to verify the prediction performance of the model, thus obtaining the trained E-FuzzyKAN-based cooling system energy consumption prediction model.

[0043] Step 4: Collect relevant operating data of the data center cooling system in real time through different sensors, input the real-time operating data into the trained E-FuzzyKAN prediction model for prediction, and output the final energy consumption prediction result of the cooling system.

[0044] In one embodiment, in step 1, different sensors are deployed in the computer room. Based on the energy transfer logic of server heat dissipation and cooling system cooling, the operating data that has a significant impact on the energy consumption of the cooling system is screened out. To ensure the quality of the dataset, the data preprocessing includes: removal of abnormal historical time series data, filling in missing values, and data normalization.

[0045] In one embodiment, in step 2, a synergistically enhanced prediction framework is formed by organically integrating the structured logical expression capability of the fuzzy inference system with the efficient nonlinear learning characteristics of the KAN neural network, thereby achieving rapid approximation of the complex dynamic characteristics of the cooling system's energy consumption. Its built-in enhancement layer can further learn the local interaction relationships between fuzzy rules through sliding convolution operations on the input data using shared convolution kernels, thereby improving the accuracy of the cooling system's energy consumption prediction. The specific method includes the following steps:

[0046] Step 21: Use the input sequence X as the input variable of the network, where each node represents an input variable. Traditional FNNs combine fuzzy inference with MLPs, leveraging the uncertainty handling capabilities of fuzzy systems and the nonlinear mapping capabilities of MLPs. However, with increasing data complexity, the limitations of MLPs in training efficiency and generalization ability have become increasingly apparent. KAN, as an improvement on MLPs, is based on the Kolmogorov-Arnold theorem. KANs do not have linear weight parameters; all weight parameters are replaced by univariate functions parameterized as spline functions. While MLP neurons always have fixed activation functions, in KANs, weights are learnable activation functions on the edges. Each edge in a KAN represents a univariate function parameterized as a spline curve, allowing for dynamic and fine-grained adjustment based on data. The fuzzification layer transforms the precise values ​​from the input layer into fuzzy values. Each variable is mapped to multiple fuzzy sets, and this layer uses Gaussian membership functions to calculate the membership degree of each input variable to each fuzzy set. Fuzzy sets can effectively capture the uncertainty caused by error fluctuations in sensor-acquired data, thereby enhancing the system's adaptability and robustness to complex real-world data center scenarios. The specific formula is as follows:

[0047] (1)

[0048] in, The first The input number The mean and standard deviation of a fuzzy set. For the number of input variables, The number of fuzzy sets, For parameterized activation functions, This represents the degree of membership.

[0049] Step 22: This fuzzy rule layer implements fuzzy inference through fuzzy rules. Each node represents the premise part of the fuzzy rule and performs an AND fuzzy logic operation. Each edge is a learnable activation function. The output of this layer is the activation strength of the rule, which is obtained by multiplying the membership degrees of each premise part. The specific formula is as follows:

[0050] (2)

[0051] in, For the first The activation strength of a fuzzy rule It is a fuzzy rule number.

[0052] Step 23: This normalization layer normalizes the output of the fuzzy rule layer so that the sum of the activation intensities of all rules is one. The specific formula is as follows:

[0053] (3)

[0054] in, This is the normalized matrix.

[0055] Step 24: After normalizing the activation intensity, reshape it into a structured tensor representing the response of each feature fuzzy rule. This enhancement layer consists of two one-dimensional convolutional and pooling layers. Using convolutional layers allows the learning of local interactions between fuzzy rules by leveraging shared convolutional kernels. This can capture useful common activations and suppress redundant or conflicting rules. The introduction of the enhancement layer forms a deep perceptual network, which has stronger non-linear expressive power compared to direct linear mapping. The specific formula is as follows:

[0056] (4)

[0057] in, For fuzzy rule enhancement matrix, For two one-dimensional convolution operations, This is a pooling operation.

[0058] Step 25: This output layer calculates the weighted sum of the outputs of the enhancement layer, and outputs the predicted cooling system energy consumption result. The specific formula is as follows:

[0059] (5)

[0060] in, For linear weights, For bias, This is a predicted value for the energy consumption of the cooling system.

[0061] Example:

[0062] This embodiment uses a sample set of energy consumption data from a private data center to predict the energy consumption of the cooling system based on E-FuzzyKAN in this application. Table 1 compares the evaluation metrics of this invention and other advanced models on this energy consumption dataset. The comparison with other advanced models shows the effectiveness of the energy consumption prediction model of this invention. This invention is superior to other models in both mean square error and mean absolute error.

[0063] Table 1 Evaluation metrics of this invention and other models

[0064] .

[0065] Figure 1The overall process of this method is as follows: collecting multi-source data and further analyzing the collected data to obtain effective data and then creating a data sample set. The training set and test set are divided in an 8:2 ratio. The data is input into the built cooling system energy consumption prediction model for training and parameter tuning. Finally, the predicted energy consumption results are output. Based on the results, the cooling system adjustment parameters are further analyzed to enable the cooling system to adapt to the computer room temperature and reduce energy consumption.

[0066] Figure 2 This diagram illustrates the operational logic of heat exchange between the cooling system and the server. The server generates a large amount of heat under different workloads. This hot air returns to the cooling system through the hot aisle, and the cool air from the cooling system is sent to the area around the server through the cold aisle to lower the temperature. This diagram reflects the logical relationship from server energy consumption fluctuations to air supply parameter adjustment to cooling energy consumption response.

[0067] Figure 3 This is a network diagram integrating KAN network and fuzzy system. The network diagram has six layers from input to output. The design concept comes from the fact that fuzzy system can accurately represent the fuzzy correlation between server energy consumption fluctuations, air supply temperature and air supply volume. It transforms the empirical fuzzy rules in data center operation and maintenance into computable mathematical logic, which can effectively solve the problem that traditional models have difficulty in handling fuzzy coupling relationships between variables. KAN network can efficiently explore the deep nonlinear mapping law between cooling system energy consumption and multi-dimensional input variables with its strong fitting ability, making up for the shortcomings of fuzzy system in fitting complex time series data. The organic integration of the two significantly improves the accuracy and stability of cooling system energy consumption prediction.

[0068] Figure 4 This is a comparison chart of predictions from different models on this energy consumption data sample set. Based on the comparison of predicted values ​​and actual values, the comparison models selected are advanced algorithms that fuse different neural networks and fuzzy systems, which can more intuitively show the advantages of this invention in terms of prediction accuracy compared to other advanced models.

Claims

1. A method for predicting the energy consumption of a cooling system based on E-FuzzyKAN, characterized in that: Includes the following steps: step :1: Collect operational data affecting the energy consumption of the data center cooling system, including server energy consumption, supply air temperature, and supply air volume; preprocess the operational data to obtain a model data sample set; Step 2: Construct a cooling system energy consumption prediction model based on E-FuzzyKAN; Step 3: Divide the data sample set into a training set and a test set. Use the training set to train the E-FuzzyKAN prediction model built in Step 2. Use the test set to verify the prediction performance of the model and obtain the trained E-FuzzyKAN-based cooling system energy consumption prediction model. Step 4: Input the real-time collected operating data into the trained E-FuzzyKAN prediction model for prediction, and output the final energy consumption prediction result of the cooling system.

2. The energy consumption prediction method for the cooling system of E-FuzzyKAN as described in claim 1, characterized in that: In step 1, Different sensors are deployed in the computer room, and based on the energy transfer logic of server heat dissipation and cooling system refrigeration, operational data that has a significant impact on the energy consumption of the cooling system is selected.

3. The method for predicting the cooling system energy consumption of E-FuzzyKAN as described in claim 2, characterized in that: Preprocessing of the running data includes: removal of outlier data from historical time series, filling in missing values, and data normalization.

4. The cooling system energy consumption prediction method based on E-FuzzyKAN as described in claim 1, characterized in that: In step 2: By organically integrating the structured logical expression capability of the fuzzy reasoning system with the efficient nonlinear learning characteristics of the KAN neural network, a synergistically enhanced prediction framework is formed.

5. The method for predicting the cooling system energy consumption of E-FuzzyKAN as described in claim 4, characterized in that: The specific method for forming a synergistically enhanced prediction framework is as follows: Step 21: Use the input sequence X as the input variable of the network, where each node represents an input variable. The specific formula is as follows: (1) in, The first The input number The mean and standard deviation of a fuzzy set. For the number of input variables, The number of fuzzy sets, For parameterized activation functions, Membership degree; Step 22: The fuzzy rule layer implements fuzzy inference through fuzzy rules. The output is the activation strength of the fuzzy rule, which is obtained by multiplying the membership degrees of each premise part. The specific formula is as follows: (2) in, For the first The activation strength of a fuzzy rule It is a fuzzy rule number; Step 23: The normalization layer normalizes the output of the fuzzy rule layer so that the sum of the activation intensities of all fuzzy rules is one. The specific formula is as follows: (3) in, This is a normalized matrix; Step 24: After normalizing the activation intensity, reshape it into a structured tensor representing the fuzzy rule response of each feature, as shown in the following formula: (4) Among them, among them, For fuzzy rule enhancement matrix, For two one-dimensional convolution operations, This is a pooling operation; Step 25: The output layer calculates the weighted sum of the outputs of the enhancement layer, outputting the predicted cooling system energy consumption result. The specific formula is as follows: (5) in, For linear weights, For bias, This is a predicted value for the energy consumption of the cooling system.

Images