Cold load short-term prediction method based on improved convolutional gated cycle unit network
By improving the convolutional gated cycle unit network and data preprocessing technology, the accuracy of the data center cooling load prediction is solved, the cooling output is matched with the required cooling load, reducing energy waste and improving the refrigeration system Energy efficiency.
Patent Information
- Application Number
- CN202410981919.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-22
- Publication Date
- 2024-11-15
AI Technical Summary
The prior art is difficult to accurately predict changes in the cooling load of the data center, resulting in the disagreement output of the refrigeration system's cooling capacity does not match the cooling load required at the end, resulting in waste of energy.
The improved convolutional gating recurrent unit network is adopted, and the data preprocessing and neural network model is built, and the historical data of cold load is calculated in combination with the principle of thermodynamics. The wavelet denoising and linear normalization methods are used to build the convolutional gating recurrent unit network model for Cooling load forecast.
Accurate short-term prediction of cooling loads is achieved, energy waste is avoided, scientific guidance on the operation of the refrigeration system, and the energy efficiency of the refrigeration system is improved.
Smart Images

Figure CN118964930A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of cold load change data analysis and prediction, and specifically discloses a cold load short-term prediction method based on an improved convolutional gated cyclic unit network. Background Art
[0002] Data centers are a type of large-scale building with high energy consumption and rapid growth, accounting for a growing proportion of the country's energy consumption year by year. The energy consumption of data centers is relatively concentrated. In addition to IT equipment, auxiliary facilities such as refrigeration systems account for a high proportion of energy consumption. IT equipment needs to run 24 hours a day throughout the year, which will consume a lot of energy and be accompanied by very high internal loads. In order to ensure the stable operation of IT equipment, the refrigeration system will also run all day and consume a lot of energy to ensure its normal operation. Energy-saving transformation of data centers usually focuses on improving the efficiency of the refrigeration system and reducing the loss of cooling capacity. Improving the energy efficiency of the refrigeration system is of great significance to the energy saving of data centers.
[0003] The data center cooling system includes chillers, cooling water systems, chilled water systems, cooling towers, and cooling media. In order to ensure the stable operation of the system, the control strategy is relatively conservative, and the frequency conversion function of most equipment is not enabled, resulting in a mismatch between the cooling output of the cooling system and the required cooling load at the end, causing energy waste. The key to saving energy lies in the prediction of the cooling load at the end, and the accuracy of the prediction will also affect the subsequent optimization and scheduling of the chillers.
[0004] Models for solving building energy consumption prediction can be divided into two categories: physical models and data-driven models. Among them, the application of physical models to simulate building energy consumption models requires the collection of detailed building information based on the use of physical equations. Some detailed building parameters cannot be obtained, resulting in poor prediction results. Data-driven models perform statistical analysis and machine learning based on historical data, and are widely used because of their good learning and generalization capabilities. The law of change of building cooling load has the characteristics of randomness, periodicity, nonlinearity and non-stationarity. When ordinary neural networks are used for prediction, it is easy to fall into local optimality and overfitting occurs, thus affecting the prediction results. Recurrent neural networks can effectively avoid this problem, but are prone to the "gradient disappearance" problem. Summary of the invention
[0005] The purpose of the present invention is to provide a short-term prediction method for cooling load based on an improved convolutional gated cyclic unit network. The method can more accurately predict and regulate the cooling capacity of the refrigeration system, avoid the mismatch between the cooling output and the required cooling load at the terminal, and cause energy waste. It can also scientifically guide the operation of the refrigeration system, which is also the inevitable direction for the current optimization and improvement of energy efficiency of my country's refrigeration system.
[0006] The objective of the present invention is achieved through the following technical solutions:
[0007] A short-term cooling load prediction method based on an improved convolutional gated recurrent unit network is proposed. A neural network model is established through data. After the model is trained, the cooling load at future moments can be predicted more accurately. The method specifically includes the following steps:
[0008] 1) Collect a large amount of accurate data required for the experiment;
[0009] 2) Calculate the required cooling load by thermodynamic principles and obtain historical cooling load data;
[0010] 3) De-noising the signal and pre-processing the data by improving the wavelet denoising method, and finally obtaining the required experimental data;
[0011] 4) After scaling the data using a linear normalization method, the processed data set is divided into a training set and a test set;
[0012] 5) Build a convolutional gated recurrent unit network model, set corresponding parameters, and obtain a prediction model;
[0013] 6) Based on the prediction model, predict the future cooling load.
[0014] The cooling load prediction method based on the convolutional gated cyclic unit network, the step 1) collects a large amount of accurate data required for the experiment, including outdoor dry-bulb temperature data and wet-bulb temperature data.
[0015] In the cooling load prediction method based on convolutional gated recurrent unit network, the cooling load historical data is obtained in step 2):
[0016] The cooling load is calculated by direct measurement of the cooling load based on the medium specific heat c, the chilled water supply and return water temperature difference ΔT and the mass flow rate m, as shown in (1):
[0017] Q=c×m×ΔT(1)
[0018] In the cooling load prediction method based on the convolution gated recurrent unit network, the step 3) performs data preprocessing to obtain the data required for the experiment; in the wavelet decomposition, by setting a suitable threshold and making a simple judgment, the main components of the signal and the wavelet coefficients with larger amplitudes are retained; the coefficients with smaller amplitudes are usually considered to be noise, and are suppressed, and the retained signal is reconstructed to achieve the purpose of removing noise; the key to the effective noise removal of this method lies in the selection of wavelet basis and threshold; Symlets are selected as wavelet basis functions according to the characteristics of the signal; compared with the hard threshold, the soft threshold function can better retain the detail information, and can also avoid over-smoothing, can better reflect the change of cooling load, and is more in line with the actual situation; the soft threshold function is as follows (2):
[0019]
[0020] Where: X is the original wavelet coefficient, η is the processed wavelet coefficient, and λ is the threshold value; when the absolute value of the wavelet coefficient is less than λ, it is set to 0, otherwise, it is subtracted from λ and takes a positive value;
[0021] Since the data collection process is affected by equipment and human factors, it is easy to cause data missing, data anomalies, etc., which will affect the results of data analysis. In order to obtain better prediction results, the data is preprocessed; linear interpolation is used to fill the missing values in the data set, such as (3); outliers in the data are processed through quartile anomaly detection;
[0022]
[0023] In the cooling load prediction method based on the convolutional gated recurrent unit network, after the data is scaled using a linear normalization method in step 4), data of different orders of magnitude will satisfy the standard normal distribution, and the normalization method is as follows (4); the processed data set is divided into a training set and a test set at a ratio of 9:1, wherein the training set is used to train the neural network to achieve higher accuracy, and the test set is used to determine the accuracy of the neural network prediction;
[0024]
[0025] Among them, X' represents the normalized value, X represents the original value, and X min and X max Represent the minimum and maximum values in the sample set respectively.
[0026] In the cooling load prediction method based on the convolutional gated recurrent unit network, the prediction model in step 5) is set as a multi-input single-output system; the parameters input into the model include the cooling load value at the previous moment, the cooling load value at the current moment, the outdoor temperature at the next moment, and the wet-bulb temperature at the next moment, and the output is the cooling load value at the next moment;
[0027] A heat load prediction model based on a temporal convolutional network is built, and the corresponding parameters are set to obtain the prediction model. The convolutional gated recurrent unit network is mainly composed of an input layer, a convolutional layer, a BN layer, a GRU layer, an Attention layer, a fully connected layer, and an output layer. GRU is an improvement on the traditional recurrent neural network and is widely used in sequence modeling and natural language processing. The input layer receives historical load data and external factor data such as weather, and performs data preprocessing. The CNN model is used to extract the spatial features of time series data. The BN layer processes the input of each hidden layer to deal with the problem of internal covariate shift. The GRU model is used to model the dependency between time series data. The Attention layer is introduced to optimize the weights of the GRU layer output to strengthen the influence of related characteristics. The output layer outputs the final prediction structure. The final prediction result is obtained by combining the attention output with the output of the GRU model. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] Figure 1 This is a flow chart of load forecasting in the present invention;
[0029] Figure 2 This is the GRU structure diagram of the present invention;
[0030] Figure 3 This is a schematic diagram of Attention of the present invention;
[0031] Figure 4 This is a diagram of the cooling load prediction model of the convolutional gated recurrent unit network of the present invention. DETAILED DESCRIPTION
[0032] In order to clearly explain the purpose, technical details and effective application of the present invention and facilitate the understanding and implementation of ordinary technicians in the field, the following will be further described in detail in conjunction with the embodiments of the present invention and the accompanying drawings. Obviously, the implementation examples described here are only used to illustrate and explain the present invention and are not used to limit the present invention.
[0033] The present invention provides a method for short-term prediction of cooling load based on an improved convolutional gated cyclic unit network, comprising:
[0034] 1) Collect a large amount of accurate data required for the experiment;
[0035] 2) Calculate the required cooling load by thermodynamic principles and obtain historical cooling load data;
[0036] 3) De-noising the signal and pre-processing the data by improving the wavelet denoising method, and finally obtaining the required experimental data;
[0037] 4) After scaling the data using a linear normalization method, the processed data set is divided into a training set and a test set;
[0038] 5) Build a convolutional gated recurrent unit network model, set corresponding parameters, and obtain a prediction model;
[0039] 6) Based on the prediction model, predict the future cooling load.
[0040] The data collected in step 1 mainly include outdoor dry-bulb temperature data and wet-bulb temperature data.
[0041] Step 2 uses the direct measurement method of cooling load to calculate the cooling load based on the medium specific heat c, the temperature difference ΔT between the supply and return water of the chilled water, and the mass flow rate m, as shown in (1):
[0042] Q=c×m×ΔT(1)
[0043] Step 3 In the wavelet decomposition, by setting a suitable threshold and making a simple judgment, the main components of the signal and the wavelet coefficients with large amplitude are retained; the coefficients with small amplitude are usually considered to be noise and are suppressed, and the retained signal is reconstructed to achieve the purpose of removing noise. The key to this method's ability to effectively remove noise lies in the selection of wavelet basis and threshold. Symlets are selected as wavelet basis functions based on the characteristics of the signal. Compared with the hard threshold, the soft threshold function can better retain detail information, while also avoiding over-smoothing, and can better reflect the changes in cooling load, which is more in line with actual conditions. The soft threshold function is shown in (2).
[0044]
[0045] Where: X is the original wavelet coefficient, η is the processed wavelet coefficient, and λ is the threshold value; when the absolute value of the wavelet coefficient is less than λ, it is set to 0, otherwise, it is subtracted from λ and takes a positive value;
[0046] Since the data collection process is affected by equipment and human factors, it is easy to cause data missing, data anomalies, etc., which will affect the results of data analysis. In order to obtain better prediction results, the data is preprocessed; linear interpolation is used to fill the missing values in the data set, such as (3); outliers in the data are processed through quartile anomaly detection;
[0047]
[0048] After scaling the data using the linear normalization method described in step 4, data of different orders of magnitude will satisfy the standard normal distribution, and the normalization method is as follows (4). The processed data set is divided into a training set and a test set at a ratio of 9:1, where the training set is used to train the neural network to achieve higher accuracy, and the test set is used to determine the accuracy of the neural network prediction.
[0049]
[0050] Among them, X′ represents the normalized value, X represents the original value, and X min and X max Represent the minimum and maximum values in the sample set respectively.
[0051] Step 5: The cooling load prediction model is set as a multi-input single-output system. The parameters input into the model include the cooling load value at the previous moment, the cooling load value at the current moment, the outdoor temperature at the next moment, and the wet bulb temperature at the next moment, and the output is the cooling load value at the next moment.
[0052] A cooling load prediction model based on a convolutional gated recurrent unit network is constructed. The convolution is regarded as a sliding filter on a cooling load sequence, which is called a one-dimensional convolution, as shown in Equation (5).
[0053]
[0054] Where: is the i-th feature in the sequence; is the jth time step in feature i; ω is the weight assigned to each time step in each feature; b is the bias. Apply one-dimensional convolution to generate a set of high-order features from the sequence of current, adjacent past and future states. This set of high-order features contains important information in the sequence in a short period of time, which can help the GRU model better understand and predict time series data. To deal with the problem of internal covariate shift, batch normalization is used to process the input of each hidden layer. Compared with LSTM, GRU uses an update gate to replace the input gate and forget gate of LSTM, which simplifies the control of forgetting factor and state update, thereby forming a model with smaller computation and faster training speed. The GRU model can predict future outputs based on previous inputs and states. More complex models are constructed by stacking GRU layers to improve the prediction effect of the model. The Attention mechanism is introduced to act on the CNN-GRU model to give adaptive weights to its output. The Attention-CNN-GRU hybrid model can accurately predict the cooling load of the power station.
[0055] Step 6 predicts the future cooling load based on the optimized prediction model.
[0056] Compared with the prior art, the technical solution provided by the present invention has the following technical effects and advantages at least: The present invention provides a cooling load prediction method based on an improved convolutional gated cyclic unit network, which performs short-term prediction of the cooling load of the power station in view of the problem of energy waste caused by the mismatch between the cooling output of the refrigeration system and the required cooling load at the terminal. Since the accuracy of the data is affected by equipment and human factors, improved wavelet denoising is used to perform data denoising, and the historical cooling load is calculated by the principle of thermodynamics, and accurate cooling load data is obtained after data preprocessing. The processed cooling load, outdoor temperature, wet-bulb temperature, etc. are input into the CNN-GRU model to predict the short-term cooling load of the power station. The CNN-GRU model has great potential in short-term cooling load prediction, and the present invention can provide a reliable data source for the on-demand allocation of cooling capacity of the refrigeration system.
[0057] Example
[0058] Reference Manual Attached Figure 1 The present invention comprises the following steps:
[0059] Step 1: The collected data mainly includes outdoor dry-bulb temperature data and wet-bulb temperature data;
[0060] Step 2: Use the direct measurement method of cooling load to calculate the cooling load based on the medium specific heat c, the temperature difference ΔT between the supply and return water of chilled water, and the mass flow rate m;
[0061] Step 3: In wavelet decomposition, by setting appropriate thresholds and making simple judgments, the main components of the signal and the wavelet coefficients with larger amplitudes are retained; the coefficients with smaller amplitudes are usually considered to be noise and are suppressed, and the retained signal is reconstructed to achieve the purpose of removing noise. The key to this method's ability to effectively remove noise lies in the selection of wavelet basis and thresholds. Symlets are selected as wavelet basis functions based on the characteristics of the signal. Compared with the hard threshold, the soft threshold function can better retain detail information, while also avoiding over-smoothing, and can better reflect changes in cooling loads, which is more in line with actual conditions;
[0062] Step 4: After scaling the data using the linear normalization method, the processed dataset is divided into a training set and a test set;
[0063] Step 5: Build a convolutional gated recurrent unit network model, set the corresponding parameters, and obtain the prediction model. The convolutional gated recurrent unit network is mainly composed of an input layer, a convolutional layer, a BN layer, a GRU layer, an Attention layer, a fully connected layer, and an output layer. The GRU structure is described in the Appendix of the manual. Figure 2 , which is an improvement on the traditional recurrent neural network and is widely used in fields such as sequence modeling and natural language processing. Figure 3The input layer receives historical load data and external factors such as weather data, and performs data preprocessing. The CNN model is used to extract the spatial features of time series data. The BN layer processes the input of each hidden layer to deal with the internal covariate offset problem. The GRU model is used to model the dependencies between time series data. The Attention layer optimizes the weights of the GRU layer output to strengthen the influence of related characteristics. The output layer outputs the final prediction structure. The final prediction result can be obtained by combining the attention output with the output of the GRU model.
[0064] Step 6: The cooling load prediction model is set as a multi-input single-output system. The parameters input into the model include the cooling load value at the previous moment, the cooling load value at the current moment, the outdoor temperature at the next moment, and the wet-bulb temperature at the next moment, and the output is the cooling load value at the next moment.
[0065] Reference Manual for Cooling Load Prediction Model Based on Convolutional Gated Recurrent Unit Network Figure 4 Based on experience and manual tuning, the hyperparameter settings of the convolutional gated recurrent unit network model are determined, with the number of one-dimensional convolution kernels being 20, the kernel time domain length being 4, the kernel stride being 2, and the initial learning rate of the GRU layer being 0.05. Based on the prediction model, the future cooling load is predicted.
[0066] The above specific implementation modes are used to explain the present invention rather than to limit the present invention. Any modification and change made to the present invention within the purpose of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.
Claims
1. A method for short-term cooling load prediction based on an improved convolutional gated recurrent unit network, characterized in that: The method establishes a neural network model through data, and after the model is trained, it can more accurately predict the cooling load at a future moment, and specifically includes the following steps: 1) Collect a large amount of accurate data required for the experiment; 2) Calculate the required cooling load by thermodynamic principles and obtain historical cooling load data; 3) De-noising the signal and pre-processing the data by improving the wavelet denoising method, and finally obtaining the required experimental data; 4) After scaling the data using a linear normalization method, the processed data set is divided into a training set and a test set; 5) Build a convolutional gated recurrent unit network model, set corresponding parameters, and obtain a prediction model; 6) Based on the prediction model, predict the future cooling load.
2. A cooling load prediction method based on a convolutional gated recurrent unit network according to claim 1, characterized in that: The step 1) collects a large amount of accurate data required for the experiment, including outdoor dry-bulb temperature data and wet-bulb temperature data.
3. The cooling load prediction method based on convolutional gated recurrent unit network according to claim 1 is characterized in that: In the cooling load prediction method based on convolutional gated recurrent unit network, the cooling load historical data is obtained in step 2): The cooling load is calculated by direct measurement of the cooling load based on the medium specific heat c, the chilled water supply and return water temperature difference ΔT and the mass flow rate m, as shown in (1): Q = c × m × ΔT (1).
4. The cooling load prediction method based on convolutional gated recurrent unit network according to claim 1 is characterized in that: In the cooling load prediction method based on the convolution gated recurrent unit network, the step 3) performs data preprocessing to obtain the data required for the experiment; in the wavelet decomposition, by setting a suitable threshold and making a simple judgment, the main components of the signal and the wavelet coefficients with larger amplitudes are retained; the coefficients with smaller amplitudes are usually considered to be noise, and are suppressed, and the retained signal is reconstructed to achieve the purpose of removing noise; the key to the effective noise removal of this method lies in the selection of wavelet basis and threshold; Symlets are selected as wavelet basis functions according to the characteristics of the signal; compared with the hard threshold, the soft threshold function can better retain the detail information, and can also avoid over-smoothing, can better reflect the change of cooling load, and is more in line with the actual situation; the soft threshold function is as follows (2): Where: X is the original wavelet coefficient, η is the processed wavelet coefficient, and λ is the threshold value; when the absolute value of the wavelet coefficient is less than λ, it is set to 0, otherwise, it is subtracted from λ and takes a positive value; Since the data collection process is affected by equipment and human factors, it is easy to cause data missing, data anomalies, etc., which will affect the results of data analysis. In order to obtain better prediction results, the data is preprocessed; linear interpolation is used to fill the missing values in the data set, such as (3); outliers in the data are processed through quartile anomaly detection; 5. The cooling load prediction method based on convolutional gated recurrent unit network according to claim 1 is characterized in that: In the cooling load prediction method based on the convolutional gated recurrent unit network, after the data is scaled using a linear normalization method in step 4), data of different orders of magnitude will satisfy the standard normal distribution, and the normalization method is as follows (4); the processed data set is divided into a training set and a test set at a ratio of 9:1, wherein the training set is used to train the neural network to achieve higher accuracy, and the test set is used to determine the accuracy of the neural network prediction; Among them, X' represents the normalized value, X represents the original value, and X min and X max Represent the minimum and maximum values in the sample set respectively.
6. The cooling load prediction method based on convolutional gated recurrent unit network according to claim 1 is characterized in that: In the cooling load prediction method based on the convolutional gated recurrent unit network, the prediction model in step 5) is set as a multi-input single-output system; the parameters input into the model include the cooling load value at the previous moment, the cooling load value at the current moment, the outdoor temperature at the next moment, and the wet-bulb temperature at the next moment, and the output is the cooling load value at the next moment; A heat load prediction model based on a temporal convolutional network is built, and the corresponding parameters are set to obtain the prediction model. The convolutional gated recurrent unit network is mainly composed of an input layer, a convolutional layer, a BN layer, a GRU layer, an Attention layer, a fully connected layer, and an output layer. GRU is an improvement on the traditional recurrent neural network and is widely used in sequence modeling and natural language processing. The input layer receives historical load data and external factor data such as weather, and performs data preprocessing. The CNN model is used to extract the spatial features of time series data. The BN layer processes the input of each hidden layer to deal with the problem of internal covariate shift. The GRU model is used to model the dependency between time series data. The Attention layer is introduced to optimize the weights of the GRU layer output to strengthen the influence of related characteristics. The output layer outputs the final prediction structure. The final prediction result is obtained by combining the attention output with the output of the GRU model.
Citation Information
Patent Citations
Short-term wind speed prediction method based on two-stage decomposition, LSTM and AT
CN111950759A
Intelligent building non-intrusive load decomposition method based on parallel connection network
CN115392318A
Heating ventilation air conditioner cooling capacity prediction method and system and related equipment
CN116578844A
Office building cooling load data calibration and prediction method based on multi-feature fusion
CN116882264A
Method and system for improving energy efficiency of large water-cooling central air-conditioning system
CN117553404A
Cited By
Cold load prediction method based on weight clustering and Pyraformer time sequence model
CN120596964A
A cold load prediction method based on weight clustering and a pyraformer time series model
CN120596964B
