Improved CNN (Convolutional Neural Network) air-cooled refrigerator energy consumption online estimation algorithm combined with dynamic feature extension
By dynamically extending and improving the CNN model, the problems of low accuracy and computational complexity in refrigeration energy consumption estimation are solved, achieving efficient and accurate online energy consumption prediction, which is suitable for embedded systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ROBOTICS RESEARCH CENTER OF YUYAO CITY
- Filing Date
- 2025-12-17
- Publication Date
- 2026-05-08
AI Technical Summary
Existing methods for estimating energy consumption in freezers fail to adequately consider dynamic characteristics, resulting in low model accuracy, computational complexity, and high costs, making it difficult to meet the needs of high-efficiency energy saving and control strategy optimization.
An improved CNN model combining dynamic feature extension is adopted. By collecting the operating status data of the freezer, preprocessing and standardizing it, extracting features using a convolutional neural network, constructing a lightweight improved CNN, estimating energy consumption, and achieving online energy consumption prediction through quantization and Adam optimization.
It improves the accuracy and adaptability of refrigerator energy consumption prediction, meets the real-time operation requirements of low-power embedded systems, and has great application potential.
Smart Images
Figure CN121996905A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of artificial intelligence technology, specifically relating to an improved CNN-based online energy consumption estimation algorithm for air-cooled refrigerators that incorporates dynamic feature extensions. Background Technology
[0002] With the widespread application of air-cooled freezers in home appliances and commercial sectors, how to efficiently estimate the operating energy consumption of freezers has become an important research direction for overall energy saving and control strategy optimization.
[0003] Most existing energy consumption estimation methods rely on traditional mathematical statistical models or sensor-based data analysis. In addition, some traditional neural network methods, such as multilayer perceptrons and simple convolutional neural networks, while improving accuracy, still suffer from the following technical problems:
[0004] (1) The mathematical statistical model does not fully consider dynamic characteristics, cannot effectively capture the time dependence between the working state of the freezer and the external environment, and the accuracy of the model is difficult to guarantee.
[0005] (2) Using sensors to calculate energy consumption is costly.
[0006] (3) Existing energy consumption prediction methods have relatively complex models and require a large amount of computation.
[0007] Therefore, there is an urgent need for a new method that combines historical data of the freezer with current data by introducing dynamic feature extension technology, thereby enhancing the model's adaptability to changes in the freezer's operating status and external environment, and thus improving the accuracy of energy consumption prediction. Summary of the Invention
[0008] To address the aforementioned technical problems in the existing technology, the purpose of this invention is to improve the accuracy of energy consumption prediction for freezers, and the technical solution is as follows:
[0009] An improved CNN-based online energy consumption estimation algorithm for air-cooled refrigerators, incorporating dynamic feature extensions, includes the following steps:
[0010] Step 1: Collect the operating status data of the freezer to form an operating status vector, preprocess the operating status vector to form continuous and stable time series data, and standardize the time series data;
[0011] Step 2: Use a convolutional neural network to extract features from the standardized time series data, and use a sliding time window to concatenate the standardized data of the current time and the historical time to form extended input features, calculate the magnitude of changes in the running state, and determine the length of the time window;
[0012] Step 3: Construct a lightweight improved CNN, perform one-dimensional convolution on the time series data to extract local temporal features, perform convolution and pooling to form high-level features, and perform regression through fully connected layers to output energy consumption estimates;
[0013] Step 4: Quantize the model parameters and intermediate features, converting the floating-point data into integer data with a lower bit width;
[0014] Step 5: Construct training samples using historical operating data, use mean squared error as the loss function to calculate the difference loss between the energy consumption estimate and the actual energy consumption, and use Adam to optimize the parameters. If the preset conditions are met, it is determined to be converged.
[0015] Step 6: Solidify the converged model parameters and deploy them to the freezer controller. Run the process cyclically online and output the results in real time. As an online energy consumption estimation result.
[0016] Furthermore, in step 1, the operating status data includes: compressor operating frequency. Evaporator fan speed Condenser fan speed Air temperature inside the cabinet and ambient temperature ;
[0017] The expression for the running state vector is:
[0018]
[0019] in, Indicates the first The state vector is formed by sampling times.
[0020] Furthermore, in step 1, the preprocessing operations include: data synchronization, outlier handling, missing value compensation, and smoothing operations.
[0021] Furthermore, in step 1, the standardization involves converting the freezer operating status parameters with different physical meanings and dimensions into time-series input data of a unified scale. The standardization processing calculation expression is as follows:
[0022]
[0023] in, This represents the standardized sensor data. This represents the mean of the data. The standard deviation of the data is represented by the number of digits; then, in the digits... The standardized running state vector expression obtained at each sampling time is:
[0024]
[0025] in, This represents the standardized operating parameters of the compressor. This represents the standardized operating parameters of the evaporator fan. This represents the standardized operating parameters of the condenser fan. This represents the standardized cabinet temperature parameters. This represents the standardized ambient temperature parameters.
[0026] Furthermore, in step 2, the calculation process for feature extraction is as follows:
[0027] For any input time series, feature extraction is performed, and the expression for calculating the output feature after convolution is as follows:
[0028]
[0029] in, For input signal, For convolution kernel weights, The length of the convolution kernel;
[0030] The calculation expression for the extended input feature is:
[0031]
[0032] in, This represents the expanded input features. This represents the standardized operating state vector at the current moment. Indicates the time step. Indicates the first Each sampling time;
[0033] The expression for calculating the magnitude of the change in the operating state is:
[0034]
[0035] in, This is the standardized operating state vector at the current moment;
[0036] The expression for calculating the length of the time window is:
[0037]
[0038] in, , These are the lower and upper limits of the window length, respectively. For adjustment coefficients, This indicates that the results will be limited to this range. This indicates rounding down to the nearest integer.
[0039] Furthermore, in step 3, the calculation process for the energy consumption estimate is as follows:
[0040] (1) For time series data, one-dimensional convolution is used to extract local dependencies in the time dimension. The feature output calculation expression for each convolutional kernel is as follows:
[0041]
[0042] in, Indicates the time position of the input features The value of , Indicates the first The convolutional kernel at the _th ... The weight of each position, Indicates the kernel length. Indicates the bias term. This represents the feature output after convolution;
[0043] (2) The convolution output is input into the pooling layer for downsampling, and the pooling output expression is:
[0044]
[0045] in, Indicates the pooling window size. This represents the output features after pooling;
[0046] (3) Flatten the features after multi-layer convolution and pooling and input them into the fully connected layer for fusion, and output the energy consumption estimate. The expression for calculating the energy consumption estimate is as follows:
[0047]
[0048] in, This represents the high-level feature vector obtained by flattening after convolution and pooling. Represents the weight matrix. For bias terms, Indicates the first The estimated energy consumption at each sampling time.
[0049] Furthermore, the number of convolutional layers is 2 to 4, and the number of convolutional kernels in each convolutional layer is 16, with a kernel length of 3.
[0050] The pooling window size of the pooling layer is 2, and the number of neurons in the fully connected layer is 16.
[0051] Furthermore, in step 4, the quantization process converts the floating-point data into integer data with a lower bit width. The quantized data is converted using the following formula, expressed as:
[0052]
[0053] in, Represents floating-point data. Indicates the quantization step size. This represents the quantized integer value.
[0054] Furthermore, in step 4, the quantization process uses different quantization methods for the convolutional layer and the fully connected layer. The specific strategies are: (1) 8-bit quantization is used for the convolutional layer; (2) 4-bit quantization is used for the fully connected layer.
[0055] Furthermore, in step 5, the calculation expression for the difference loss between the predicted value and the actual value is as follows:
[0056]
[0057] in, Indicates the number of samples. Represents the actual value. Indicates the predicted value;
[0058] The parameter optimization using Adam is expressed as follows:
[0059]
[0060] in, and These represent the first and second moment estimates of the gradient, respectively. Indicates the learning rate. To prevent small constants from being divided by zero.
[0061] Beneficial effects: (1) By introducing dynamic feature extension, improving the CNN structure and combining lightweight processing strategies, the accuracy of cold cabinet energy consumption prediction is improved. (2) It has high adaptability and accuracy in the face of environmental changes and real-time computing requirements. (3) It meets the real-time operation requirements of low-power embedded systems and has great application potential. Attached Figure Description
[0062] Figure 1 The flowchart of the improved CNN online energy consumption estimation algorithm for air-cooled refrigerators combined with dynamic feature extension is shown below.
[0063] Figure 2 This is a diagram illustrating the online energy consumption estimation effect of the present invention. Detailed Implementation
[0064] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.
[0065] like Figure 1 As shown, the improved CNN-based online energy consumption estimation algorithm for air-cooled refrigerators, which incorporates dynamic feature extension, includes the following steps:
[0066] Step 1: Data Acquisition and Preprocessing
[0067] During the operation of the freezer, the sampling period is used. Real-time acquisition of freezer operating status data, including: (1) compressor operating parameters (1) indicates the compressor operating frequency; (2) fan operating parameters , indicating the evaporator fan speed; (3) Fan operating parameters (4) Temperature parameters (5) Temperature parameters This indicates the temperature environment. The operating status data is then used to construct the... The running state vector at each sampling time:
[0068]
[0069] in, Indicates the first The state vector is formed by sampling times.
[0070] A series of preprocessing operations, including data synchronization, outlier handling, missing value compensation, and smoothing, are performed on the running state vector to eliminate differences in the dimensions of different physical quantities, reduce the interference of noise and missing values on the model, and form continuous and stable time series data. Since the various running parameters have significant differences in dimensions and numerical ranges, each physical quantity is standardized to have a mean of 0 and a variance of 1. For any running parameter... Standardize;
[0071]
[0072] in, This represents the standardized sensor data. This represents the mean of the data. The standard deviation of the data.
[0073] Then in the The standardized running state vector expression obtained at each sampling time is:
[0074]
[0075] in, This represents the standardized operating parameters of the compressor. This represents the standardized operating parameters of the evaporator fan. This represents the standardized operating parameters of the condenser fan. This represents the standardized cabinet temperature parameters. This represents the standardized ambient temperature parameters.
[0076] Through data acquisition and standardization, the operating status parameters of refrigerated display cases with different physical meanings and dimensions are converted into time series input data of a unified scale, providing a standardized data foundation for subsequent dynamic feature expansion and improvement of feature extraction of convolutional neural networks.
[0077] Step 2: Feature Extraction and Dynamic Feature Expansion
[0078] After completing the data acquisition and standardization process in step 1, the standardized freezer operating status data is used as the input to the neural network model. Through feature extraction and dynamic feature expansion design, the model's ability to understand the time series characteristics of the freezer operating status is enhanced.
[0079] To address the significant temporal correlation in freezer operation data, a convolutional neural network is employed to extract features from the time series data, thereby learning the local temporal features between adjacent time steps.
[0080] For any input feature sequence, the convolution operation can be represented as:
[0081]
[0082] in, For input signal, For convolution kernel weights, The length of the convolution kernel. This represents the output features after convolution.
[0083] To enable the time window length to adaptively adjust to changes in the operating conditions of the freezer, this invention calculates an operating state change index based on the degree of change in a standardized operating state vector, and dynamically determines the window length accordingly. When the operating status change index increases, it indicates that the operating conditions of the freezer are changing more drastically. To improve the response to rapid changes, the time window length should be reduced. When the operating status change index decreases, it indicates that the operating status is relatively stable. To enhance the learning ability of long-term dependent characteristics, the time window length should be increased.
[0084] A sliding time window is used to concatenate standardized data from the current moment and historical moments to form extended input features, assuming a time step of [value missing]. Then in the first At each sampling time, the expanded input features are:
[0085]
[0086] in, This represents the expanded input features. This represents the standardized operating state vector at the current moment.
[0087] Define the range of changes in running status:
[0088]
[0089] in, This is the standardized operating state vector at the current moment.
[0090] Set the window length as a monotonically decreasing function of the magnitude of change; that is, the greater the change, the shorter the window.
[0091]
[0092] in, , These are the lower and upper limits of the window length, respectively. For adjustment coefficients, This indicates that the results will be limited to this range. This indicates rounding down to the nearest integer.
[0093] Through convolutional feature extraction and dynamic feature expansion processing, the model can automatically learn key features from the original operating status data. These key features are used to characterize the coordinated operation status of the compressor and fan, the temperature change trend, and the comprehensive impact of environmental factors on the energy consumption of the freezer, and serve as inputs for subsequent improvement of the convolutional neural network model for energy consumption estimation.
[0094] Step 3: Improvement of Convolutional Neural Networks (CNN)
[0095] We construct a lightweight and improved CNN model suitable for time series data to extract time series features from dynamically expanded running state inputs and output energy consumption estimates.
[0096] For time series data, one-dimensional convolution is used to extract local temporal features in the time dimension. For the... A convolution kernel, whose convolution output can be represented as:
[0097]
[0098] in, Indicates the time position of the input features The value of , Indicates the first The convolutional kernel at the _th ... The weight of each position, Indicates the kernel length. Indicates the bias term. This represents the feature output after convolution.
[0099] To reduce feature dimensionality and suppress noise, the output of the convolutional layer is downsampled by a pooling layer. Taking max pooling as an example, its pooling output is:
[0100]
[0101] in, Indicates the pooling window size. This represents the output characteristics after pooling.
[0102] The features obtained from multi-layer convolution and pooling are flattened and then fed into a fully connected layer for fusion, outputting an energy consumption estimate. The output of the fully connected layer is:
[0103]
[0104] in, This represents the high-level feature vector obtained by flattening after convolution and pooling. Represents the weight matrix. For bias terms, Indicates the first The estimated energy consumption at each sampling time.
[0105] By using a smaller kernel length This reduces the computational cost of a single convolutional layer; by stacking multiple convolutional layers, the network can gain a larger receptive field of view while keeping the computational cost under control, thereby learning more complex temporal features.
[0106] When predicting the energy consumption of refrigerators, the number of convolutional layers is set to 2 to 4, the number of convolutional kernels in each convolutional layer is set to 16, the kernel length is set to 3, the pooling window size of the pooling layer is set to 2, and the number of neurons in the fully connected layer is set to 16.
[0107] Step 4: Lightweight Data Processing
[0108] To ensure the algorithm runs efficiently in embedded systems, lightweight design and quantization optimization techniques are employed to reduce computational and storage overhead. Floating-point data is converted into lower-bit-width integer data. The quantized data is converted using the following formula:
[0109]
[0110] in, Represents floating-point data. Indicates the quantization step size. This represents the quantized integer value.
[0111] Different quantization methods are used in convolutional and fully connected layers. Convolutional layers use 8-bit quantization to reduce memory consumption and computational overhead, while fully connected layers use 4-bit quantization to further improve the computational efficiency of the model.
[0112] Step 5: Train the model
[0113] Training samples are constructed using historical operational data, with the input being... The supervised output is the actual energy consumption, and the mean squared error (MSE) is used as the loss function. After each training round or several training rounds, the difference loss between the predicted energy consumption and the actual energy consumption is calculated:
[0114]
[0115] in, Indicates the number of samples. Represents the actual value. This represents the predicted value.
[0116] The Adam optimization algorithm is used, and the formula is as follows:
[0117]
[0118] in, and These represent the first and second moment estimates of the gradient, respectively. Indicates the learning rate. To prevent small constants from being divided by zero.
[0119] Convergence is determined when the preset conditions are met.
[0120] Step 6: Algorithm Deployment
[0121] The parameters of the trained and converged model are solidified and deployed to the freezer controller, which executes them cyclically during online operation and outputs results in real time. As an online energy consumption estimation result.
[0122] The online energy consumption estimation performance of the prediction algorithm is as follows: Figure 2 As shown in the figure, the predicted power and the actual power of the algorithm are close throughout the compressor's operating cycle, proving the effectiveness of the online prediction algorithm for the freezer.
Claims
1. An improved CNN-based online energy consumption estimation algorithm for air-cooled refrigerators, incorporating dynamic feature extensions, characterized in that: Includes the following steps: Step 1: Collect the operating status data of the freezer to form an operating status vector, preprocess the operating status vector to form continuous and stable time series data, and standardize the time series data; Step 2: Use a convolutional neural network to extract features from the standardized time series data, and use a sliding time window to concatenate the standardized data of the current time and the historical time to form extended input features, calculate the magnitude of changes in the running state, and determine the length of the time window; Step 3: Construct a lightweight improved CNN, perform one-dimensional convolution on the time series data to extract local temporal features, perform convolution and pooling to form high-level features, and perform regression through fully connected layers to output energy consumption estimates; Step 4: Quantize the model parameters and intermediate features, converting the floating-point data into integer data with a lower bit width; Step 5: Construct training samples using historical operating data, use mean squared error as the loss function to calculate the difference loss between the energy consumption estimate and the actual energy consumption, and use Adam to optimize the parameters. If the preset conditions are met, it is determined to be converged. Step 6: Solidify the converged model parameters and deploy them to the freezer controller. Run the process cyclically online and output the results in real time. As an online energy consumption estimation result.
2. The improved CNN-based online energy consumption estimation algorithm for air-cooled refrigerators combined with dynamic feature extension as described in claim 1, characterized in that, In step 1, the operating status data includes: compressor operating frequency. Evaporator fan speed Condenser fan speed Air temperature inside the cabinet and ambient temperature ; The expression for the running state vector is: in, Indicates the first The state vector is formed by sampling times.
3. The improved CNN-based online energy consumption estimation algorithm for air-cooled refrigerators, incorporating dynamic feature extensions, as described in claim 1, is characterized in that... In step 1, the preprocessing operations include: data synchronization, outlier handling, missing value compensation, and smoothing operations.
4. The improved CNN-based online energy consumption estimation algorithm for air-cooled refrigerators, incorporating dynamic feature extension, as described in claim 1, is characterized in that... In step 1, the standardization process converts the operating status parameters of the freezer with different physical meanings and dimensions into time-series input data of a unified scale. The standardization calculation expression is as follows: in, This represents the standardized sensor data. This represents the mean of the data. The standard deviation of the data is represented by the number of digits; then, in the digits... The standardized running state vector expression obtained at each sampling time is: in, This represents the standardized operating parameters of the compressor. This represents the standardized operating parameters of the evaporator fan. This represents the standardized operating parameters of the condenser fan. This represents the standardized cabinet temperature parameters. This represents the standardized ambient temperature parameters.
5. The improved CNN-based online energy consumption estimation algorithm for air-cooled refrigerators combined with dynamic feature extension as described in claim 1, characterized in that, In step 2, the calculation process for feature extraction is as follows: For any input time series, feature extraction is performed, and the expression for calculating the output feature after convolution is as follows: in, For input signal, For convolution kernel weights, The length of the convolution kernel; The calculation expression for the extended input feature is: in, This represents the expanded input features. This represents the standardized operating state vector at the current moment. Indicates the time step. Indicates the first Each sampling time; The expression for calculating the magnitude of the change in the operating state is: in, This is the standardized operating state vector at the current moment; The expression for calculating the length of the time window is: in, , These are the lower and upper limits of the window length, respectively. For adjustment coefficients, This indicates that the results will be limited to this range. This indicates rounding down to the nearest integer.
6. The improved CNN-based online energy consumption estimation algorithm for air-cooled freezers, incorporating dynamic feature extension, as described in claim 1, is characterized in that... In step 3, the calculation process for the energy consumption estimate is as follows: (1) For time series data, one-dimensional convolution is used to extract local dependencies in the time dimension. The feature output calculation expression for each convolutional kernel is as follows: in, Indicates the time position of the input features The value of , Indicates the first The convolutional kernel at the _th ... The weight of each position, Indicates the kernel length. Indicates the bias term. This represents the feature output after convolution; (2) The convolution output is input into the pooling layer for downsampling, and the pooling output expression is: in, Indicates the pooling window size. This represents the output features after pooling; (3) Flatten the features after multi-layer convolution and pooling and input them into the fully connected layer for fusion, and output the energy consumption estimate. The expression for calculating the energy consumption estimate is as follows: in, This represents the high-level feature vector obtained by flattening after convolution and pooling. Represents the weight matrix. For bias terms, Indicates the first The estimated energy consumption at each sampling time.
7. The improved CNN-based online energy consumption estimation algorithm for air-cooled refrigerators, incorporating dynamic feature extension, as described in claim 6, is characterized in that... The number of convolutional layers is 2 to 4, and each convolutional layer has 16 convolutional kernels with a kernel length of 3. The pooling window size of the pooling layer is 2, and the number of neurons in the fully connected layer is 16.
8. The improved CNN-based online energy consumption estimation algorithm for air-cooled refrigerators combined with dynamic feature extension as described in claim 1, characterized in that, In step 4, the quantization process converts the floating-point data into integer data with a lower bit width. The quantized data is converted using the following formula, expressed as: in, Represents floating-point data. Indicates the quantization step size. This represents the quantized integer value.
9. The improved CNN-based online energy consumption estimation algorithm for air-cooled refrigerators, incorporating dynamic feature extension, as described in claim 1, is characterized in that... In step 4, the quantization process uses different quantization methods for the convolutional layer and the fully connected layer. The specific strategies are: (1) 8-bit quantization is used for the convolutional layer; (2) 4-bit quantization is used for the fully connected layer.
10. The improved CNN-based online energy consumption estimation algorithm for air-cooled refrigerators combined with dynamic feature extension as described in claim 1, characterized in that, In step 5, the difference loss between the predicted value and the actual value is calculated using the following expression: in, Indicates the number of samples. Represents the true value. Indicates the predicted value; The parameter optimization using Adam is expressed as follows: in, and These represent the first and second moment estimates of the gradient, respectively. Indicates the learning rate. To prevent small constants from being divided by zero.