A physical constraint data-driven modeling method for variable frequency start-stop of cascade refrigeration system

Abstract

This invention discloses a data-driven modeling method for variable frequency start-stop of a cascade refrigeration system based on physical constraints. First, it collects multi-dimensional data on the physical, control, switching, and time states of the cascade refrigeration system throughout its entire lifecycle. Transient and steady-state data are separated, with transient sampling prioritized and densified. Switching and physical components are aligned and filtered. Then, after standardization, discrete state embedding, and time vectorization, interactive coupling features such as thermal hysteresis and thermal coupling temperature difference are constructed to form a time-series training set. Next, a dual-stream hybrid network is used to extract independent features for high and low temperatures, which are then fused with interactive coupling features and active power features. These features are then concatenated with the corresponding measured residuals to predict state increments, outputting a prediction vector. Finally, a composite loss function incorporating data fitting error, cascaded thermal coupling constraints, and discrete start-stop logic constraints is used to complete model training. The model constructed in this invention can effectively improve the accuracy of time-series prediction during the start-stop phase, ensure physical consistency of the results, and is suitable for engineering applications of dynamic control during variable frequency start-stop.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent control and complex thermal equipment modeling technology, and in particular to a physical constraint data-driven method for modeling the start-stop of a cascade refrigeration system using frequency conversion. Background Technology

[0002] Cascade refrigeration systems consist of two independent refrigeration cycles—a high-temperature stage and a low-temperature stage—coupled by an intermediate condenser-evaporator. They are core equipment for obtaining cryogenic environments ranging from -40°C to -80°C and below. However, the high- and low-temperature cycles in cascade refrigeration systems are mutually restrictive, especially under variable load and start-up / shutdown conditions where refrigerant migration is complex, resulting in drastic dynamic fluctuations in the internal temperature field. These systems exhibit strong nonlinear and thermodynamic coupling characteristics, making it impossible to fully characterize the system's transient operating state and achieve high-precision dynamic control by relying solely on conventional sensors to collect single-point steady-state parameters. To achieve high-precision dynamic control, existing research has attempted to construct start-up / shutdown models for cascade refrigeration systems, but current model construction methods still have significant limitations.

[0003] Existing modeling methods mainly fall into two categories: mechanistic modeling and purely data-driven modeling. Mechanistic modeling relies on complex differential equations, making it difficult to accurately describe the nonlinear flow characteristics and oil film establishment process during the start-up and shutdown phases of variable frequency compressors. It also incurs high computational loads and is difficult to identify parameters. On the other hand, purely data-driven modeling methods often treat all inputs as continuous variables, lacking the embedding of physical constraints on discrete start-up and shutdown events. This results in start-up and shutdown transient predictions violating thermodynamic laws and failing to accurately capture the phenomena of thermal inertia hysteresis during shutdown and instantaneous cooling during startup. Summary of the Invention

[0004] The present invention aims to solve the above-mentioned technical problems existing in the prior art by providing a physical constraint data-driven method for modeling the start-stop of a cascade refrigeration system.

[0005] The technical solution of this invention is: a physical constraint data-driven method for modeling the start-stop of a cascade refrigeration system, characterized by performing the following steps sequentially: Step 1. Data Acquisition and Preprocessing The system collects full-cycle multidimensional data reflecting the core thermodynamic processes of the cascade refrigeration system, including continuous physical components, continuous control components, discrete logic state components, and time dynamic components. The full-cycle multidimensional data is divided into transient and steady-state segments, and the transient segments are prioritized and sampled in a more intensive manner. The discrete logic state components and continuous physical components are aligned according to time sequence. The continuous physical and control components are filtered and cleaned. Step 2. Training Dataset Construction The continuous physical components and continuous control components are standardized; the discrete logic state components are encoded; the time dynamic components are vectorized; thermal hysteresis features and thermal coupling temperature difference features are constructed; the continuous time series data formed by the processing and construction results are divided into fixed-length learning samples and divided into high-temperature level subsets, low-temperature level subsets, interactive coupling subsets and active power subsets. Step 3. Extract high-temperature and low-temperature features based on a parallel long short-term memory network. The high-temperature subset and the low-temperature subset are respectively input into the corresponding channels of the parallel long short-term memory network to extract the high-temperature feature and the low-temperature feature; Step 4. Fuse features and predict The high-temperature level features, low-temperature level features, interactive coupling subsets, and active power subsets are concatenated to obtain fused features; the fused features are predicted based on the prediction head to obtain state increments; then the state increments are added to the residuals of the standardized continuous physical components to output the predicted features. Step 5. Model training with multiple physics constraints Model training is completed based on predictive features and a composite loss function that includes data fitting error, cascaded thermal coupling constraints, and discrete start-stop logic constraints.

[0006] Preferably, the continuous physical components are the low-temperature stage condensation temperature, the low-temperature stage evaporation temperature, the high-temperature stage evaporation temperature, the high-temperature stage condensation temperature, the ambient temperature, the temperature of the object being cooled, and the total active power; the continuous control components are the low-temperature stage compressor speed and the high-temperature stage compressor speed; the discrete logic state components are the power-on state and the power-off state; and the time dynamic component is the duration of the current state.

[0007] The transient segment includes the start-up segment, the shutdown segment, and the speed regulation segment. The start-up segment is the period from when the compressor starts up until the power stabilizes. The shutdown segment is the period from when the shutdown command is issued until the high and low pressures are balanced. The speed regulation segment is the period in which the inverter frequency undergoes a step change. The steady-state segment is the period in which the system temperature fluctuation is within the allowable temperature difference range. Preferably, the encoding of discrete logic state components is performed through a learnable embedding layer, which encodes the power-on / off state variables. Mapped to a Dense vectors of dimension ; The vectorization of the time variable refers to the duration of the current state. A nonlinear transformation is performed to obtain the duration. The first after time vectorization encoding Dimensional components : ; in, Indicates duration The first after time vectorization encoding Dimensional components; Indicates the first Frequency parameters of time-dimension encoding; Indicates the first Phase parameters of 3D time encoding; The dimension index representing the periodic encoded item; Represents the highest dimension of the periodic coding term; The thermal hysteresis feature is constructed by creating a virtual power feature that decays exponentially according to the following formula at the instant the system receives a shutdown command: ; in, It is characterized by thermal hysteresis; From power-on to power-off time The standardized final power; It is the current sampling time after the power-off moment; The system thermal time constant; The thermal coupling temperature difference characteristic of the structure is calculated according to the following formula: cascade temperature difference characteristic. : ; in, This refers to the standardized low-temperature stage condensation temperature. This refers to the standardized high-temperature evaporation temperature.

[0008] The preferred parallel long short-term memory network includes a high-temperature flow channel and a low-temperature flow channel with consistent structure; each channel adopts a vertically stacked L-layer bidirectional long short-term memory module, and each layer extracts forward and reverse temporal features respectively. The forward and reverse temporal features are spliced ​​and fused to obtain the output features of the current layer. After L-layer progressive feature extraction, high-temperature level features and low-temperature level features are output respectively.

[0009] The preferred composite loss function is as follows: ; in, This represents the data fitting error; For cascaded thermal coupling constraints; For discrete start / stop logic constraints; and These are adaptive hyperparameters used to balance the strength of physical constraints; ; in, It is a predictive feature; These are standardized measured values; This is a preset threshold parameter; t It is the index of the time series sample; N is the total number of time series samples; ; Where RELU(·) is a linear rectified function; It predicts the evaporation temperature of the high-temperature stage; It is the predicted condensation temperature of the low-temperature stage; This is the preset minimum temperature difference parameter; ; Where RELU(·) is a linear rectified function; To predict power; It is the timing step index of the shutdown decay phase; This is the total number of timing steps during the shutdown decay phase; ; in, Indicates duration as The theoretical upper bound of the system power attenuation; The standardized final power from power-on to power-off time; It is the duration of the current state; It is the attenuation coefficient; This is the bias parameter for power compensation of waste heat attenuation.

[0010] Compared with the prior art, the present invention has the following advantages: (1) The original historical data is divided into steady-state segment, start-up segment, shutdown segment and speed regulation segment according to the operation logic. Oversampling is performed on the start-up segment, shutdown segment and speed regulation segment, and undersampling is performed on the steady-state segment. This solves the problem of redundancy of steady-state samples and scarcity of transient samples in industrial data, forces the neural network to fully learn the nonlinear fluid dynamic characteristics during start-up and shutdown, avoids the model from falling into steady-state overfitting, and improves the prediction accuracy of start-up impact and shutdown lag.

[0011] (2) By introducing a discrete state learnable embedding layer, the shortcomings of traditional continuous-time models in being unable to identify discrete events such as power on and off are effectively solved, enabling the model to accurately distinguish the essential differences between transient and steady-state operation in the time dimension, and improving the prediction network's physical understanding of the power on and off logic. When the power is on, the mapping vector automatically learns weights that contain physical meanings such as high power expectation and rapid pressure build-up. When the power is off, the mapping vector automatically learns weights that contain physical meanings such as residual heat dissipation.

[0012] (3) By constructing shutdown thermal hysteresis characteristics and cascade thermo-coupling characteristics, the final power and high and low temperature stage heat exchange boundary are locked at the moment of shutdown command. During the shutdown period, virtual power characteristics decaying exponentially are constructed, forcing the model to remember the energy level before shutdown and the physical relationship of cascade heat exchange. This accurately predicts the temperature inertia change and residual cold release process after shutdown, avoids the non-physical phenomenon of the predicted value instantly returning to zero, and solves the prediction distortion problems caused by refrigerant migration, residual cold release on metal walls and cascade heat exchange mismatch after shutdown.

[0013] (4) By constructing a parallel long short-term memory network, the low-temperature flow channel and the high-temperature flow channel respectively use a stacked bidirectional long short-term memory network to extract their respective temporal features, and use past and future information to complete the prediction. The topology of the physical system is explicitly mapped to the hierarchical structure of the neural network, decoupling the dynamic characteristics of the high and low temperature stage cycles of the cascade refrigeration system, avoiding feature interference caused by mixing all variables as input, allowing the network to learn the dynamic laws of the two high and low temperature cycle loops respectively, and improving the modeling accuracy of the cascade coupling characteristics of the cascade refrigeration system.

[0014] (5) By embedding cascaded temperature difference penalty and long-term power decay boundary constraints in the loss function, on the one hand, the model is forced to comply with the thermodynamic law of heat transfer from the low temperature stage to the high temperature stage in the cascade refrigeration system, and the invalid predictions that violate the heat transfer logic are completely corrected; on the other hand, the temperature lag and power slow-down characteristics brought about by residual cold release are accurately reproduced, and the non-physical deviation of the prediction value instantaneously returning to zero is eliminated from the root; ensuring that the prediction results under all operating conditions are consistent with the thermodynamic response law of the real system. Attached Figure Description

[0015] Figure 1 This is an overall architecture diagram of an embodiment of the present invention.

[0016] Figure 2 This is a flowchart of step 2 in an embodiment of the present invention.

[0017] Figure 3 This is a flowchart of steps 3 and 4 in an embodiment of the present invention.

[0018] Figure 4 This is a waveform comparison diagram of the predicted values ​​and actual sensor values ​​of key state variables in an embodiment of the present invention. Detailed Implementation

[0019] This invention provides a physical constraint data-driven method for modeling the variable frequency start-stop of a cascade refrigeration system, such as... Figure 1 The following steps will be performed sequentially as shown: Step 1. Data Acquisition and Preprocessing This step aims to collect multivariate coupled time-series data from the complex industrial operating environment of cascade refrigeration systems. Through a physically-guided signal processing flow, it filters out electromagnetic interference and fluid pulsation noise caused by variable frequency drives, constructing a high-quality dataset that accurately reflects the start-up and shutdown transients and cascade thermodynamic processes. The specific steps are as follows: Step 1.1 Collect full-cycle multidimensional data of the continuous physical components, continuous control components, discrete logic state components and time dynamic components of the cascade refrigeration system that reflect the core thermodynamic process. In order to capture the rapid dynamic changes at the moment of compressor start-up and shutdown, the sampling frequency is set to be no less than 1Hz. The continuous physical components are the low-temperature stage condensing temperature, low-temperature stage evaporating temperature, high-temperature stage evaporating temperature, high-temperature stage condensing temperature, ambient temperature, temperature of the object being cooled (refrigeration cabinet), and total active power; the continuous control components are the low-temperature stage compressor speed and the high-temperature stage compressor speed; the discrete logic state component is the power-on state. and power off It can directly read the start / stop command bits of the controller and define... , where 1 represents the system being on or running, and 0 represents the system being off or in standby mode; the time dynamic component is the duration of the current state, which is the cumulative time since the last state transition (such as from 0 to 1, or from 1 to 0).

[0020] Step 1.2 divides the full-cycle multidimensional data into transient and steady-state segments. Prioritizing and encrypting the transient segments, the segmented sampling is completed. This forces the set neural network to fully learn the nonlinear fluid dynamic characteristics during the start-up and shutdown process, rather than simply memorizing the linear mapping under steady-state conditions. The transient segment includes the start-up segment, the shutdown segment, and the speed regulation segment. The start-up segment is the period from the issuance of the compressor start command to the establishment of stable power. The shutdown segment is the period from the issuance of the shutdown command to the balance of high and low pressure. The speed regulation segment is the period in which the frequency of the inverter undergoes a step change. The steady-state segment is the period in which the system temperature fluctuation is within the allowable temperature difference range, such as the system temperature fluctuation within ±0.3℃ to ±1.0℃. Specifically, the sampling can be carried out in a sample ratio of 3:3:1:3 for the start-up segment, the shutdown segment, the speed regulation segment, and the steady-state segment.

[0021] Step 1.3 Align the discrete logic state components with the continuous physical components according to their timing. Because there is a time discrepancy between the controller logic signal and the sensor's physical response, and because the physical process after shutdown is continuous, the raw logic data needs to be preprocessed specifically, using the power sensor reading jump as a reference. The signal undergoes microsecond-level time calibration to eliminate tag misalignment caused by communication delays. Specifically, the sampling sequence of the power sensor is traversed to find the microsecond-level timestamp when the power jumps from the steady-state mean value beyond a threshold. Tp Repositioning timestamp of signal level transition Ts Calculate the time difference ΔT = Tp - Ts Shift the time axis of the switching logic signals as a whole. ΔT This means aligning discrete logic state components with continuous physical components in a time sequence.

[0022] Step 1.4 Filters and cleans the continuous physical components and continuous control components; Given the prevalent high-frequency electromagnetic noise and occasional outliers in industrial sensor data, and to avoid introducing additional computational burden through overly complex signal processing, this step employs a mature and robust cascaded filtering strategy to filter continuous physical and control components. This improves data quality while preserving start-stop dynamic characteristics. The specific filtering operations for continuous physical and control components are as follows: First, outlier handling based on the 3-σ criterion is performed to address abnormal jumps caused by sensor malfunctions or packet loss. Within a sliding time window, the mean of the data is calculated. and standard deviation If a certain sampling point satisfy If an outlier is identified, it is removed to eliminate large "outliers"; then a second-order Butterworth low-pass filter is used to smooth the continuous physical variables. Set cutoff frequency This frequency should be much lower than the system's thermal response frequency, but higher than the main dynamic frequency during start-up and shutdown. Its transfer function modulus square response is: ; in, This represents the frequency response of a second-order Butterworth low-pass filter at angular frequency ω; ω represents the angular frequency of the input signal. Indicates the cutoff frequency The corresponding angular frequency; n This represents the order of the Butterworth low-pass filter. This zero-phase filtering process can effectively filter out high-frequency glitches superimposed on the real signal and prevent phase lag caused by traditional filtering, ensuring that the waveform does not shift in time at the moment of start-up and stop. Cleaning the continuous physical components and continuous control components is based on the fundamental laws of thermodynamics (such as evaporation temperature being lower than condensation temperature). This process removes abnormal samples that violate physical principles due to sensor malfunctions, ensuring the physical consistency of the input data.

[0023] Step 2. Training dataset construction, execute as follows Figure 2The following steps are shown: Step 2.1 Standardize the continuous physical components and continuous control components; To eliminate the dimensional differences between different physical quantities (such as temperature -80℃ and speed 4000rpm), the Z-Score standardization method is used to standardize the continuous physical components and continuous control components.

[0024] Step 2.2 Encodes the discrete logic state components; To address the issue that directly inputting the discrete variable of power on / off state into a neural network yields poor results, this step utilizes a learnable embedding layer to process the power on / off state variable... Mapped to a Dense vectors of dimension ;when (Upon startup), dense vectors automatically learn weights that incorporate physical meanings such as high power expectation and rapid pressure build-up; when When the system is powered off, dense vectors automatically learn weights that include physical meanings such as residual heat dissipation.

[0025] Step 2.3 Vectorize the time state components; To capture the impact of action time on system state, a time vectorization coding technique is introduced, which encodes the duration of the current state. A nonlinear transformation is performed to obtain the duration. The first after time vectorization encoding Dimensional components : ; in, Indicates duration The first after time vectorization encoding Dimensional components; Indicates the first Frequency parameters of time-dimension encoding; Indicates the first Phase parameters of 3D time encoding; The dimension index representing the periodic encoded item; The highest dimension of the periodic encoding term is a preset positive integer; Through this encoding, the model can accurately distinguish start-stop transients (small). Value) and steady-state operation (large) The essential difference in value over time.

[0026] Step 2.4 Construct thermal hysteresis characteristics and thermally coupled temperature difference characteristics; To address the spurious power or delayed temperature recovery caused by refrigerant migration and residual cold release from metal walls after shutdown, a thermal hysteresis feature is designed. This feature, constructed at the instant the system receives the shutdown command, exhibits a virtual power characteristic that decays exponentially during the subsequent shutdown period, according to the following formula: ; in, It is characterized by thermal hysteresis; From power-on to power-off time The standardized final power; It is the current sampling time after the power-off moment; The system thermal time constant can be obtained by exponential fitting of historical shutdown data. This feature forces the model to remember the energy level before shutdown, thereby accurately predicting the temperature inertia change after shutdown and avoiding the non-physical phenomenon of the predicted value instantly returning to zero. To reflect the strong coupling relationship between the high-temperature stage and the low-temperature stage of the cascade system, a thermally coupled temperature difference characteristic, i.e., a cascaded temperature difference characteristic, is constructed. This feature directly reflects the heat transfer driving force of the intermediate condenser evaporator, which helps the model to quickly capture the energy transfer dynamics between the two cycles during frequency conversion regulation. The cascade temperature difference characteristics are calculated using the following formula. : ; in, This refers to the standardized low-temperature stage condensation temperature. This refers to the standardized high-temperature evaporation temperature.

[0027] Step 2.5 divides the continuous time series data formed by the processing and construction results into learning samples of fixed length, and divides them into high temperature level subset, low temperature level subset, interactive coupling subset and active power subset; The sliding time window method is used to divide continuous time series data into supervised learning samples suitable for deep learning models, that is, at time t... Construct a length of m (like m=60 The input matrix (representing the past 60 time steps) Unlike traditional inputs that only contain continuous variables, the input matrix constructed in this invention... Its elements contain information in four dimensions: ; in, For standardized continuous physical components; thermal hysteresis characteristics and cascade temperature difference characteristics ; For standardized continuous control components; dense vector ; The first time vectorized encoding Dimensional components ; Input matrix The system is divided into a high-temperature stage subset, a low-temperature stage subset, an interactive coupling subset, and an active power subset. The high-temperature stage subset includes the high-temperature stage evaporation temperature. High-temperature condensation temperature Ambient temperature High-temperature compressor speed High-temperature on / off embedded vector and the first high-temperature level time vectorized encoding Dimensional components The cryogenic stage subset includes the cryogenic stage condensation temperature. Low-temperature stage evaporation temperature Refrigeration cabinet temperature Low-temperature stage compressor speed Low-temperature stage on / off embedded vector and the first low-temperature time vectorized encoding Dimensional components The interactive coupling subset is characterized by thermal hysteresis. and thermal coupling temperature difference characteristics The active power subset is the total active power of the system. .

[0028] Step 3. Extract high-temperature and low-temperature features based on a parallel long short-term memory network. The high-temperature subset and the low-temperature subset are respectively input into the corresponding channels of the parallel long short-term memory network to extract the high-temperature feature and the low-temperature feature; The parallel long short-term memory network, such as Figure 3 As shown, it includes high-temperature flow channels and low-temperature flow channels with consistent structures. The high-temperature flow channel is designed to capture the response of the high-temperature stage to the ambient temperature and its ability to provide a condensation environment for the low-temperature stage. The low-temperature flow channel is designed to capture the thermodynamic response of the low-temperature stage and its cooling characteristics on the object being cooled (such as a freezer). Each channel adopts a vertically stacked L=2-layer bidirectional long short-term memory module. The first layer contains 128 hidden units and the second layer contains 64 hidden units to further enhance the model's ability to understand contextual information. Each layer extracts forward and reverse temporal features respectively. The forward and reverse temporal features are concatenated and fused to obtain the output features of the current layer. After L-layer progressive feature extraction, the high-temperature stage features and low-temperature stage features are output respectively. No. Layer in time step Hidden state Calculated by the following formula: ; LSTM stands for Long Short-Term Memory. It is the first The output of the layer at the current time step, and It is the first The hidden state of the layer in the previous time step; any time step The final output of this layer By splicing forward-hidden states and backward hidden state The result is represented as follows: ; This structure allows the model to utilize both past and future information when making predictions for the current time step.

[0029] Step 4. Fuse features and predict like Figure 3 As shown, the high-temperature level features, low-temperature level features, interactive coupling subsets, and active power subsets are concatenated to obtain the fused features. ; Based on prediction head MLP to fuse features Predictions are made to obtain the state increments of temperature and power. Then increment the state. and t Standardized measured values ​​of time The residual addition operation, i.e., the element-wise addition operation, outputs the predicted features. The specific meaning is as follows:

[0030] ; Step 5. Model training with multiple physics constraints Based on predictive features The loss function with multiple physical constraints is used, and the gradient descent algorithm is used to perform end-to-end training on the two-stream hybrid network to complete the model training; The specific composite loss function is as follows: ; in, This represents the data fitting error; For cascaded thermal coupling constraints; For discrete start / stop logic constraints; and These are adaptive hyperparameters used to balance the strength of physical constraints; By explicitly penalizing predictions that violate physical common sense in the composite loss function, the model is forced to learn to follow the laws of thermodynamics while using only sparse sensor data. Mean squared error is used to measure the predicted value Compared with actual sensor values To improve robustness to outliers during start-stop cycles, the differences between them are addressed by introducing a gradient in regions with larger gradients. Huber Loss : ; in, It is a predictive feature; These are the standardized measured values; This is a preset threshold parameter, which means... HuberLoss The threshold parameter that distinguishes between the quadratic error interval and the linear error interval; t It is the index of the time series sample; N is the total number of time series samples; The core characteristic of a cascade refrigeration system is that the heat dissipated by the low-temperature stage cycle must be absorbed by the high-temperature stage cycle (i.e., Although complex dynamic energy balance equations are not used, the temperature trends of the two cycles must be kept consistent, so a cascaded temperature difference penalty term is constructed; if the model predicts a reverse temperature difference, the loss of this term will increase sharply, forcing the network to correct the predicted trajectory. The cascaded temperature difference penalty term is as follows: ; Where RELU(·) is a linear rectified function; It predicts the evaporation temperature of the high-temperature stage; It is the predicted condensation temperature of the low-temperature stage; This is the preset minimum temperature difference, which is the non-negative safety margin used to ensure the heat transfer temperature difference margin. Discrete start-stop logic constraints are the core constraint design for variable frequency start-stop and thermal hysteresis characteristics. If the model predicts the power... If the physical limit of decay is exceeded (i.e., the predicted system not only does not decay but maintains high power consumption after the power is cut off), a penalty is applied to ensure that the model can simulate the phenomenon that the power and temperature do not drop to zero instantaneously after shutdown, but slowly fall back according to physical laws. The discrete start / stop logic constraints are as follows: ; Where RELU(·) is a linear rectified function; To predict power; It is the timing step index of the shutdown decay phase; It is the total number of timing steps;

[0031] in, Indicates duration as The theoretical upper bound of the system power attenuation; The standardized final power at the shutdown moment; It is the feature duration parameter; It is the attenuation coefficient; This is the bias parameter for power compensation of waste heat attenuation.

[0032] To verify the prediction accuracy of the model constructed using this invention, the model was tested, and the waveform comparison results of the predicted values ​​and actual sensor values ​​of key state variables are as follows: Figure 4 As shown in the figure. The results show that the prediction curve of the model constructed in this invention closely matches the actual measured values ​​of the system throughout the entire testing period, maintaining extremely high tracking accuracy even in the detailed fluctuations shown in the magnified area. This fully demonstrates that the modeling method provided by this invention can extract the true dynamic characteristics of the system and successfully construct a high-precision prediction model.

Claims

1. A physically constrained data-driven method for modeling the start-stop of a cascade refrigeration system, characterized in that... Perform the following steps in sequence: Step 1. Data Acquisition and Preprocessing Collect full-cycle multidimensional data of the cascade refrigeration system, including continuous physical components, continuous control components, discrete logic state components, and time dynamic components reflecting the core thermodynamic processes; divide the full-cycle multidimensional data into transient and steady-state segments, and complete segmented sampling with priority and density for transient segments; Align discrete logic state components with continuous physical components according to timing; Filtering and cleaning of continuous physical components and continuous control components; The continuous physical components are: low-temperature stage condensation temperature, low-temperature stage evaporation temperature, high-temperature stage evaporation temperature, high-temperature stage condensation temperature, ambient temperature, temperature of the object being cooled, and total active power. The continuous control components are the speed of the low-temperature stage compressor and the speed of the high-temperature stage compressor; The discrete logic state components are the power-on state and the power-off state; the time dynamic component is the duration of the current state; Step 2. Training Dataset Construction Standardize the continuous physical components and continuous control components; The discrete logic state components are encoded; the time dynamic components are vectorized; thermal hysteresis features and thermal coupling temperature difference features are constructed; the continuous time series data formed by the processing and construction results are divided into fixed-length learning samples and divided into high-temperature level subsets, low-temperature level subsets, interactive coupling subsets and active power subsets. The thermal hysteresis feature is constructed by creating a virtual power feature that decays exponentially according to the following formula at the instant the system receives a shutdown command: ； in, It is characterized by thermal hysteresis; From power-on to power-off time The standardized final power; It is the current sampling time after the power-off moment; The system thermal time constant; The thermal coupling temperature difference characteristic of the structure is calculated according to the following formula: cascade temperature difference characteristic. : ； in, The standardized low-temperature stage condensation temperature. This refers to the standardized high-temperature evaporation temperature. Step 3. Extract high-temperature and low-temperature features based on a parallel long short-term memory network. The high-temperature subset and the low-temperature subset are respectively input into the corresponding channels of the parallel long short-term memory network to extract the high-temperature feature and the low-temperature feature; Step 4. Fuse features and predict The high-temperature level features, low-temperature level features, interactive coupling subsets, and active power subsets are concatenated to obtain fused features; the fused features are predicted based on the prediction head to obtain state increments; and the state increments are then added to the residuals of the standardized continuous physical components to obtain the predicted features. Step 5. Model training with multiple physics constraints Model training is completed based on predicted features and a composite loss function; the composite loss function is as follows: ； in, This represents the data fitting error; For cascaded thermal coupling constraints; For discrete start / stop logic constraints; and These are adaptive hyperparameters used to balance the strength of physical constraints; ； in, It is a predictive feature; These are standardized measured values; This is a preset threshold parameter; t It is the index of the time series sample; N is the total number of time series samples; ； Where RELU(·) is a linear rectified function; It predicts the evaporation temperature of the high-temperature stage; It is the predicted condensation temperature of the low-temperature stage; This is the preset minimum temperature difference parameter; ； in, To predict power; It is the timing step index of the shutdown decay phase; This is the total number of timing steps during the shutdown decay phase; ； in, Indicates duration as The theoretical upper bound of the system power attenuation; The standardized final power from power-on to power-off time; It is the duration of the current state; It is the attenuation coefficient; This is the bias parameter for power compensation of waste heat attenuation.

2. The physical constraint data-driven variable frequency start-stop modeling method for cascade refrigeration systems according to claim 1, characterized in that: The transient segment includes the start-up segment, the shutdown segment, and the speed regulation segment. The start-up segment is the interval from the issuance of the compressor start-up command to the establishment of stable power. The shutdown segment is the interval from the issuance of the shutdown command to the balance of high and low pressure. The speed regulation segment is the interval in which the frequency of the inverter undergoes a step change. The steady-state segment is the interval in which the system temperature fluctuation is within the allowable temperature difference range.

3. The physical constraint data-driven variable frequency start-stop modeling method for cascade refrigeration systems according to claim 2, characterized in that: The encoding of discrete logic state components is performed through a learnable embedding layer, which encodes the power-on / off state variables. or Mapped to a Dense vectors of dimension ; The vectorization of the time dynamic components refers to the duration of the current state. A nonlinear transformation is performed to obtain the duration. The first after time vectorization encoding Dimensional components : ； in, Indicates duration The first after time vectorization encoding Dimensional components; Indicates the first Frequency parameters of time-dimension encoding; Indicates the first Phase parameters of 3D time encoding; The dimension index representing the periodic encoded item; This represents the highest dimension of the periodic coding term.

4. The physical constraint data-driven variable frequency start-stop modeling method for cascade refrigeration systems according to claim 3, characterized in that: The parallel long short-term memory network includes high-temperature flow channels and low-temperature flow channels with consistent structures. Each channel adopts a vertically stacked L-layer bidirectional long short-term memory module. Each layer extracts forward and reverse temporal features respectively. The forward and reverse temporal features are spliced ​​and fused to obtain the output features of the current layer. After L-layer progressive feature extraction, high-temperature level features and low-temperature level features are output respectively.

Images