Expolation model modeling method based on physical constraint width learning and prior knowledge

Abstract

The invention relates to a modeling method of a water chilling unit extrapolation model based on physical constraint width learning and priori knowledge, which comprises the following steps: based on equipment parameters and actual operation data, applying a width learning algorithm, establishing a width learning model, and carrying out preliminary prediction on an operation state; based on the actual operation data and the energy conservation law, calculating the mean square error of the cooling water outlet temperature of the width learning model, the mean square error of the electric power of the water chilling unit and the energy loss function of the water chilling unit, and establishing a self-defined physical loss function; an error back propagation mechanism and an Adam optimizer are utilized to optimize a self-defined physical loss function, weights and bias coefficients of a feature layer and an enhancement layer in a width learning algorithm are updated, the precision of a width learning model is improved, and a physical constraint width learning model is established; establishing a priori knowledge model; and combining the priori knowledge model with the physical constraint width learning model, and establishing an extrapolation model based on physical constraint width learning and priori knowledge.

Description

Technical Field

[0001] The invention relates to a modeling method of a chiller extrapolation model based on physical constraint width learning and prior knowledge, and belongs to the technical field of simulation modeling. Background Art

[0002] Under the background of "carbon neutrality and carbon peak", all walks of life are accelerating the construction of energy conservation and emission reduction, and more and more intelligent and refined management and control technologies are being applied to the HVAC field.

[0003] In industrial buildings, the energy consumption of air conditioning system operation accounts for 60% of the total energy consumption of the building, of which up to 70% comes from the refrigeration room. As one of the core equipment of the refrigeration room, the energy efficiency level of the chiller will directly affect the overall energy consumption of the refrigeration room. However, in the actual operation and maintenance process, due to considerations such as system stability and safety, operation and maintenance personnel are usually not allowed to frequently operate the chiller to achieve its optimal operating conditions. Therefore, in order to further improve the energy efficiency level of the chiller operation, it is necessary to establish an accurate mathematical model of the chiller to accurately predict the operating status of the chiller under different operating parameters and guide the control of the chiller.

[0004] In the prior art, mathematical models of chillers are roughly divided into empirical models and mechanism models.

[0005] The empirical model is trained by a large amount of measured data reflecting the performance characteristics of the chiller, and directly establishes the mapping relationship between the input variables and the output variables. The advantage of the empirical model is that the modeling process is relatively simple and the modeling efficiency is high; however, the empirical model requires a large amount of training data in the early stage of modeling, and when the model operation exceeds the range of the training data, the reliability of the model is poor.

[0006] The mechanism model describes and infers the operation process of the chiller based on the principles of thermodynamics, heat and mass transfer, momentum transfer, and energy conservation. The advantage of the mechanism model is that its prediction results are more consistent with the operation rules of the equipment, the prediction results are more accurate, and the interpretability is strong; however, due to the complex physical characteristics and operation mechanism of the chiller itself, the modeling is more difficult.

[0007] In addition, the above two types of models in the prior art are both static models, which are difficult to reflect the real-time operating status of the chiller.

[0008] As the core equipment of central air-conditioning refrigeration station, the performance of chillers will change under different operating conditions, which is the primary factor to be considered when formulating energy-saving strategies. The chiller model with extrapolation capability can provide simulation reference when formulating energy-saving scheduling strategies for central air-conditioning refrigeration stations, which is conducive to comparing the performance differences of different energy-saving strategies. Summary of the invention

[0009] The present invention proposes a modeling method for a chiller extrapolation model based on physical constraint width learning and prior knowledge, which aims to overcome the above-mentioned shortcomings of the prior art and achieve high-precision online prediction of the operating status of the chiller.

[0010] The technical solution of the present invention is a modeling method of a chiller extrapolation model based on physical constraint width learning and prior knowledge, comprising the following steps:

[0011] Step 1: Based on the chiller equipment parameters and actual operation data, a width learning algorithm is applied to establish a chiller width learning model to achieve a preliminary prediction of the chiller operation status;

[0012] Step 2: Based on the actual operation data of the chiller and the law of conservation of energy, calculate the mean square error of the cooling water outlet temperature of the width learning model, the mean square error of the chiller electric power, and the energy loss function of the chiller, and establish a custom physical loss function;

[0013] Step 3: Use the error back propagation mechanism and Adam optimizer to optimize the chiller’s custom physical loss function, update the weights and bias coefficients of the feature layer and enhancement layer in the width learning algorithm, improve the accuracy of the width learning model, and establish a physically constrained width learning model.

[0014] Step 4: Establish a priori knowledge model of the chiller;

[0015] Step 5: Combine the chiller prior knowledge model with the physical constraint width learning model to establish a chiller extrapolation model based on physical constraint width learning and prior knowledge;

[0016] Step 6: Build a dynamic data-driven application system for online data collection and calculation based on field sensors, edge computing servers and other equipment to achieve real-time prediction of the chiller operating status and online update of the chiller extrapolation model.

[0017] Preferably, in step 1, the input variables of the width learning model are chilled water outlet temperature, chilled water inlet temperature, cooling water inlet temperature, chilled water flow rate, and cooling water flow rate, and the output variables of the width learning model are cooling water outlet temperature and chiller electric power. The chilled water outlet temperature and cooling water inlet temperature reflect the operating conditions of the chiller; the cooling water inlet temperature, cooling water outlet temperature, and cooling water flow rate reflect the heat exchange capacity of the chiller condenser; the chilled water inlet temperature reflects the change of the terminal cooling load; the chilled water outlet temperature, chilled water inlet temperature, and chilled water flow rate reflect the cooling capacity of the chiller; and the chiller electric power reflects the working state of the chiller compressor. These parameters can reflect the operating state of the chiller under different working conditions.

[0018] Preferably, in step 2, the mean square error of the cooling water outlet temperature reflects the error between the predicted value and the actual value of the cooling water outlet temperature of the chiller under the historical operating parameters of the chiller. TCO It is expressed as:

[0019]

[0020] Where n is the number of samples; T cw,L is the actual value of cooling water outlet temperature, ℃; is the predicted value of the cooling water outlet temperature of the chiller model, ℃.

[0021] Preferably, in step 2, the mean square error of the chiller power reflects the error between the predicted value and the actual value of the chiller power under the historical operating parameters of the chiller. p It is expressed as:

[0022]

[0023] Where n is the number of samples; P is the actual value of the chiller power, kW; is the predicted value of the chiller's electrical power, kW.

[0024] Preferably, in step 2, the chiller energy loss function reflects the degree to which the model prediction value deviates from the physical constraint. energy It is expressed as:

[0025]

[0026] Where c is the specific heat capacity of chilled water, KJ / (kg·℃); G ew is the flow rate of chilled water, L / s; T ew,E is the chilled water inlet temperature, °C; T ew,L is the chilled water outlet temperature, °C; is the predicted value of the chiller power, kW; G cw is the cooling water flow rate, L / s; is the predicted value of cooling water outlet temperature, ℃; T cw,E is the cooling water inlet temperature, °C. min is the lower limit of the fluctuation range, kW; max is the upper limit of the fluctuation range, kW.

[0027] Preferably, in step 2, the self-defined physical loss function of the chiller is expressed as:

[0028] CustomLoss=MSE TCO +MSEp +Loss energy

[0029] Where CustomLoss is the physical loss of the chiller; MSE TCO MSE is the mean square error of the cooling water outlet temperature; p is the mean square error of the chiller power; Loss energy is the energy loss of the chiller.

[0030] Preferably, in step 3, the custom physical loss function of the width learning model is optimized using the error back propagation mechanism and the Adam optimizer to establish a physical constraint width learning model. The back propagation mechanism and the Adam optimizer mainly update the feature layer weights and the enhancement layer weights in the width learning network according to the custom physical loss function to minimize the custom physical loss function. The output layer weights still use the pseudo-inverse solution method.

[0031] Preferably, in step 4, the chiller prior knowledge model is expressed as:

[0032] Q ew =c·G ew ·(T ew,E -T ew,L )

[0033]

[0034] P=a0+a1·PLR+a2·PLR 2 +a3·PLR 3

[0035] COP=b0+b1·PLR+b2·PLR 2 +b3 PLR 3 +b4 PLR 4

[0036]

[0037] In the formula, Q ew is the cooling capacity of the evaporator, kW; c is the specific heat capacity of chilled water, KJ / (kg·℃); G ew is the flow rate of chilled water, L / s; T ew,E is the return water temperature of the chilled water, ℃; T ew,L is the supply temperature of chilled water, °C; PLR is the partial load rate of the chiller; Q n is the rated cooling capacity of the chiller, kW; P is the compressor power, kW; COP is the energy efficiency ratio of the chiller; a0-a3, b0-b4 are the coefficients in the performance curve; T cw,Lis the cooling water outlet temperature, ℃; G cw is the cooling water flow rate, L / s; T cw,E is the cooling water inlet temperature, ℃.

[0038] Preferably, in step 5, the prior knowledge model is integrated with the physical constraint width learning model, that is, the output of the prior knowledge model and the output of the physical constraint width learning model are weighted averaged, which is expressed as:

[0039]

[0040] Where Y KPCBLS is the output of the extrapolation model; Y K is the output of the prior knowledge model; Y PCBLS The output of the model is learned for the physical constraint width.

[0041] Preferably, in step 6, the dynamic data-driven application system collects field sensor data and transmits it to the extrapolation model modeling module to realize real-time prediction of the operating status of the chiller; and according to the error between the output result of the extrapolation model of the chiller and the actual measurement value, calculates the custom physical loss function, and optimizes the modeling process using the error back propagation mechanism and Adam optimizer to realize online correction of the accuracy of the extrapolation model.

[0042] The advantages of the present invention are as follows: the modeling method is reasonably designed, which helps to accurately predict the operating state of the chiller under other working conditions based on the actual operating data of the chiller, and ensures that the prediction results conform to physical laws. At the same time, a dynamic data-driven application system is built to realize the deployment and application of the above-mentioned chiller extrapolation model modeling method. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a schematic diagram of the flow chart of the modeling method of the present invention.

[0044] Figure 2 It is a schematic diagram of the implementation process of the modeling method of the present invention.

[0045] Figure 3 : is an experimental data graph obtained in an embodiment of the present invention, wherein (a) is a frequency distribution histogram of the chilled water outlet temperature; (b) is a frequency distribution histogram of the chilled water return temperature and the cooling water inlet temperature; (c) is a frequency distribution histogram of the chilled water flow rate and the cooling water flow rate; (d) is a box plot of the cooling water outlet temperature; (e) is a relationship graph between the compressor electric power and the partial load rate of the chiller under three working conditions; (f) is a relationship graph between the energy efficiency ratio of the chiller and its partial load rate under three working conditions.

[0046] Figure 4 It is a schematic diagram of the principle of the modeling method of the present invention.

[0047] Figure 5 Schematic diagram of energy conservation loss in an embodiment of the present invention.

[0048] Figure 6 1 is a validation diagram of the chiller extrapolation model in the embodiment of the present invention, wherein (a) is a MAE and RMSE comparison (evaluation index) diagram of the cooling water outlet temperature extrapolated by different models; (b) is a MAE and RMSE comparison (evaluation index) diagram of the chiller (compressor) electric power extrapolated by different models; (c) is a R comparison of the cooling water outlet temperature and the chiller (compressor) electric power extrapolated by different models. 2 Comparison (evaluation index) diagram. DETAILED DESCRIPTION

[0049] The present invention is further described in detail below with reference to examples and specific implementation methods.

[0050] like Figure 1 As shown, a modeling method of a chiller extrapolation model based on physical constraint width learning and prior knowledge includes the following steps:

[0051] Step 1: Based on the chiller equipment parameters and actual operation data, a width learning algorithm is applied to establish a chiller width learning model to achieve a preliminary prediction of the chiller operation status;

[0052] The input variables of the width learning algorithm are the chilled water outlet temperature, the chilled water inlet temperature, the cooling water inlet temperature, the chilled water volume flow rate, and the cooling water volume flow rate. The output variables of the width learning algorithm are the cooling water outlet temperature and the chiller electric power.

[0053] Among them, the chilled water outlet temperature and the cooling water inlet temperature reflect the operating conditions of the chiller; the cooling water inlet temperature, cooling water outlet temperature and cooling water flow rate reflect the heat exchange capacity of the chiller condenser; the chilled water inlet temperature reflects the change of the terminal cooling load; the chilled water outlet temperature, chilled water inlet temperature and chilled water flow rate reflect the cooling capacity of the chiller; the chiller electric power reflects the working status of the chiller compressor.

[0054] The above parameters can reflect the operating status of the chiller under different working conditions.

[0055] Step 2: Based on the actual operation data of the chiller and the law of conservation of energy, calculate the mean square error of the cooling water outlet temperature of the width learning model, the mean square error of the chiller electric power, and the energy loss function of the chiller, and establish a custom physical loss function;

[0056] The mean square error of the cooling water outlet temperature reflects the error between the predicted value and the actual value of the cooling water outlet temperature of the chiller under the historical operating parameters of the chiller. TCO It is expressed as:

[0057]

[0058] Where n is the number of samples; T cw,L is the actual value of cooling water outlet temperature, ℃; is the predicted value of the cooling water outlet temperature of the chiller model, ℃.

[0059] The mean square error of the chiller power reflects the error between the predicted value and the actual value of the chiller power under the historical operating parameters of the chiller. p It is expressed as:

[0060]

[0061] Where n is the number of samples; P is the actual value of the chiller power, kW; is the predicted value of the chiller's electrical power, kW.

[0062] The energy loss function of the chiller reflects the degree to which the model prediction value deviates from the physical constraints. energy It is expressed as:

[0063]

[0064] Where c is the specific heat capacity of chilled water, KJ / (kg·℃); G ew is the flow rate of chilled water, L / s; T ew,E is the chilled water inlet temperature, °C; T ew,L is the chilled water outlet temperature, °C; is the predicted value of the chiller power, kW; G cw is the cooling water flow rate, L / s; is the predicted value of cooling water outlet temperature, ℃; T cw,E is the cooling water inlet temperature, °C. min is the lower limit of the fluctuation range, kW; max is the upper limit of the fluctuation range, kW.

[0065] The custom physical loss function of the chiller is expressed as:

[0066] CustomLoss=MSE TCO +MSE p +Loss energy

[0067] Where CustomLoss is the physical loss of the chiller; MSE TCO MSE is the mean square error of the cooling water outlet temperature; p is the mean square error of the chiller power; Loss energy is the energy loss of the chiller.

[0068] Step 3: Use the error back propagation mechanism and Adam optimizer to optimize the chiller’s custom physical loss function, update the weights and bias coefficients of the feature layer and enhancement layer in the width learning algorithm, improve the accuracy of the width learning model, and establish a physically constrained width learning model.

[0069] The error back propagation mechanism and Adam optimizer are used to optimize the custom physical loss function of the width learning model and establish a physical constraint width learning model. The back propagation mechanism and Adam optimizer mainly update the feature layer weights and enhancement layer weights in the width learning network according to the custom physical loss function to minimize the custom physical loss function. The output layer weights are still solved using the pseudo-inverse solution method.

[0070] Step 4: Establish a priori knowledge model of the chiller;

[0071] The prior knowledge model of the chiller is expressed as:

[0072] Q ew =c·G ew ·(T ew,E -T ew,L )

[0073]

[0074] P=a0+a1·PLR+a2·PLR 2 +a3·PLR 3

[0075] COP=b0+b1·PLR+b2·PLR 2 +b3 PLR 3 +b4 PLR 4

[0076]

[0077] In the formula, Q ew is the cooling capacity of the evaporator, kW; c is the specific heat capacity of chilled water, KJ / (kg·℃); G ew is the flow rate of chilled water, L / s; T ew,E is the return water temperature of the chilled water, ℃; T ew,L is the supply temperature of chilled water, °C; PLR is the partial load rate of the chiller; Q nis the rated cooling capacity of the chiller, kW; P is the compressor power, kW; COP is the energy efficiency ratio of the chiller; a0-a3, b0-b4 are the coefficients in the performance curve; T cw,L is the cooling water outlet temperature, ℃; G cw is the cooling water flow rate, L / s; T cw,E is the cooling water inlet temperature, ℃.

[0078] Step 5: Combine the chiller prior knowledge model with the physical constraint width learning model to establish a chiller extrapolation model based on physical constraint width learning and prior knowledge;

[0079] The prior knowledge model is integrated with the physical constraint width learning model, that is, the output of the prior knowledge model and the output of the physical constraint width learning model are weighted averaged, which can be expressed as:

[0080]

[0081] Where Y KPCBLS is the output of the extrapolation model; Y K is the output of the prior knowledge model; Y PCBLS The output of the model is learned for the physical constraint width.

[0082] Step 6: Build a dynamic data-driven application system for online data collection and calculation based on field sensors, edge computing servers and other equipment to achieve real-time prediction of the chiller operating status and online update of the chiller extrapolation model.

[0083] The dynamic data-driven application system collects field sensor data and transmits it to the extrapolation model modeling module to achieve real-time prediction of the chiller's operating status. It calculates a custom physical loss function based on the error between the chiller's extrapolation model output and the actual measurement value, and optimizes the modeling process using the error back propagation mechanism and Adam optimizer to achieve online correction of the extrapolation model's accuracy.

[0084] Example

[0085] like Figure 2 As shown in FIG. 1 , a modeling method of a chiller extrapolation model based on physical constraint width learning and prior knowledge is shown, taking a chiller in a refrigeration room of a hospital in Xi'an as an implementation object, including the following steps:

[0086] Step 1: Collect the chiller operation data in the actual system through sensors and other equipment;

[0087] The rated cooling capacity of the chiller implemented in this embodiment is 6330kw, the rated power is 1194kw, the rated COP is 5.303, the chilled water flow is 251.6L / s, and the cooling water flow is 361.5L / s. By changing the cooling water inlet temperature and the chiller load rate, the cooling water outlet temperature and compressor power of the chiller under three different operating conditions are obtained. The experimental data are sampled every 4 minutes, and a total of 1080 sets of valid data are collected. The specific data are as follows Figure 3 As shown. Figure 3 (a) Figure 3 (b) Figure 3 From the frequency histogram of (c), it can be seen that the chilled water outlet temperature fluctuates around 7°C, the chilled water inlet temperature fluctuates between 8.8°C and 13°C, the chilled water flow rate fluctuates around 251.6L / s, and the cooling water flow rate fluctuates around 361.5L / s. Figure 3 It can be seen from the box plot of (d) that the cooling water outlet temperature fluctuates around 33°C, and the overall trend roughly follows a normal distribution. Figure 3 (e) and Figure 3 The scatter plots in (f) show the relationship between the electric power and PLR of the chiller under three operating conditions and the relationship between COP and PLR.

[0088] By analyzing the data collected in step 1, there is energy conservation loss, such as Figure 5 As shown, the energy conservation loss value fluctuates within the interval [-760, 760]. In this embodiment, it is stipulated that when the energy conservation loss value is within the interval [-760, 760], it is considered that there is no energy conservation loss, that is, the energy conservation loss value is 0; when the energy conservation loss value exceeds the above interval, the absolute value of the difference between it and the nearest boundary is used to replace the original value.

[0089] Step 2: Input the collected data into the modeling module of the chiller extrapolation model based on the physical constraint width learning and prior knowledge model, apply the width learning algorithm, establish the chiller width learning model, and calculate the mean square error of the model cooling water outlet temperature, the mean square error of the chiller electric power, and the chiller energy loss function, and establish the chiller custom physical loss function;

[0090] Mean square error (MSE) of cooling water outlet temperature TCO Specifically:

[0091]

[0092] Where n is the number of samples; T cw,L is the actual value of cooling water outlet temperature, ℃; is the predicted value of the cooling water outlet temperature of the chiller model, ℃.

[0093] Mean square error (MSE) of chiller power p Specifically:

[0094]

[0095] Where n is the number of samples; P is the actual value of the chiller power, kW; is the predicted value of the chiller's electrical power, kW.

[0096] Chiller energy loss function Loss energy Specifically:

[0097]

[0098] Where c is the specific heat capacity of chilled water, KJ / (kg·℃); G ew is the flow rate of chilled water, L / s; T ew,E is the chilled water inlet temperature, °C; T ew,L is the chilled water outlet temperature, °C; is the predicted value of the chiller power, kW; G cw is the cooling water flow rate, L / s; is the predicted value of cooling water outlet temperature, ℃; T cw,E is the cooling water inlet temperature, °C. min is the lower limit of the fluctuation range, kW; max is the upper limit of the fluctuation range, kW.

[0099] The custom physical loss function of the chiller is as follows:

[0100] CustomLoss=MSE TCO +MSE p +Loss energy

[0101] Where CustomLoss is the physical loss of the chiller; MSE TCO MSE is the mean square error of the cooling water outlet temperature; p is the mean square error of the chiller power; Loss energy is the energy loss of the chiller.

[0102] Step 3: Use the error back propagation mechanism and Adam optimizer to optimize the chiller’s custom physical loss function, update the weights and biases of the feature layer and enhancement layer in the width learning algorithm, improve the accuracy of the width learning model, and establish a physically constrained width learning model.

[0103] like Figure 4 As shown, the width learning network is optimized using the error back propagation mechanism and the Adam optimizer.

[0104] Step 4: Establish a priori knowledge model of the chiller;

[0105] The prior knowledge model of the chiller is as follows:

[0106] Q ew =c·G ew ·(T ew,E -T ew,L )

[0107]

[0108] P=a0+a1·PLR+a2·PLR 2 +a3·PLR 3

[0109] COP=b0+b1·PLR+b2·PLR 2 +b3 PLR 3 +b4 PLR 4

[0110]

[0111] In the formula, Q ew is the cooling capacity of the evaporator, kW; c is the specific heat capacity of chilled water, KJ / (kg·℃); G ew is the flow rate of chilled water, L / s; T ew,E is the return water temperature of the chilled water, ℃; T ew,L is the supply temperature of chilled water, °C; PLR is the partial load rate of the chiller; Q n is the rated cooling capacity of the chiller, kW; P is the compressor power, kW; COP is the energy efficiency ratio of the chiller; a0-a3, b0-b4 are the coefficients in the performance curve; T cw,L is the cooling water outlet temperature, ℃; G cw is the cooling water flow rate, L / s; T cw,E is the cooling water inlet temperature, ℃.

[0112] Step 5: Combine the chiller prior knowledge model with the physical constraint width learning model to establish a chiller extrapolation model based on the physical constraint width learning and prior knowledge model.

[0113] The output of the prior knowledge model and the output of the physical constraint width learning model are weighted averaged, specifically:

[0114]

[0115] Where Y KPCBLS is the output of the extrapolation model; Y K is the output of the prior knowledge model; Y PCBLSThe output of the model is learned for the physical constraint width.

[0116] In this embodiment, the pseudo code of the chiller unit extrapolation model modeling method is as follows:

[0117] Algorithm 1: Modeling method of chiller extrapolation model based on physical constraint width learning and prior knowledge model

[0118] Input: X = [T ew,L ,T ew,E ,T cw,E ,G ew ,G cw ]

[0119] Output: Y KPCBLS =[T cw,L ,P]

[0120] 1: For i = 0; i ≤ n do

[0121] 2: Randomly assign initial weights and bias W ei ,β ei ;

[0122] 3: Calculate the output Z of each feature window i =[φ(XW ei +β ei )];

[0123] 4:End

[0124] 5: Calculate the output Z of the feature layer n =[Z1,...,Z n ];

[0125] 6: Randomly assign initial weights and bias W h ,β h ;

[0126] 7: Calculate the output of the enhancement layer H = [ξ(Z n W h +β h )];

[0127] 8: Calculate the weight of the output layer Wm = [Z n |H]–Y OB ;

[0128] 9: Calculate the output Y of width learning BLS =[Z n |H]W m ;

[0129] 10: Calculate the custom physical loss function CustomLoss;

[0130] 11: When the custom physical loss function CustomLoss does not meet the requirements

[0131] 12: Back propagation;

[0132] 13: Update W using Adam optimizer ei ,β ei ,Wh,β h ;

[0133] 14: Calculate the output Z of each feature window i =[φ(XW ei +β ei )];

[0134] 15: Calculate the output Z of the feature layer n =[Z1,...,Z n ];

[0135] 16: Calculate the output of the enhancement layer H = [ξ(Z n W h +β h )];

[0136] 17: Calculate the weight of the output layer Wm = [Z n |H]–Y OB ;

[0137] 18:End

[0138] 19: Calculate the output Y of physical constraint width learning PCBLS =[Z n |H]W m ;

[0139] 20: Calculate the output Y of the prior knowledge model K ;

[0140] 21: Compute the output Y of knowledge plus physical constraint width learning KPCBLS ;

[0141] like Figure 6 As shown, the effectiveness of the chiller extrapolation model established in this embodiment is verified. By comparing the chiller extrapolation model with the conventional width learning model and the multi-layer perceptron model, the mean absolute error MAE, the root mean square error RMSE, and the determination coefficient R 2 As shown in Table 1, compared with the width learning model, the extrapolation model based on physical constraint width learning and prior knowledge model reduces the MAE index by 10.3%, the RMSE index by 14.3%, and the R2 index by 3.3% in the prediction accuracy of cooling water outlet temperature, and reduces the MAE index by 32.2%, the RMSE index by 29%, and the R2 index by 3.3% in the prediction accuracy of compressor electric power.2 Compared with the multi-layer perceptron model, the extrapolation model based on physical constraint width learning and prior knowledge model has reduced the MAE index by 58.1%, the RMSE index by 61.5%, and the R 2 The indicators were improved by 78.8%. In terms of compressor power prediction accuracy, the MAE indicator was reduced by 40.3%, the RMSE indicator was reduced by 40.9%, and the R 2 The index was improved by 1.0%. The chiller extrapolation model based on physical constraint width learning and prior knowledge model proposed by the present invention has a significant effect on improving the prediction accuracy.

[0142] Table 1 Evaluation indicators of different models

[0143]

[0144] The above is only a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A modeling method for a chiller extrapolation model based on physical constraint width learning and prior knowledge, characterized in that: The steps include: Step 1: Based on the chiller equipment parameters and actual operation data, a width learning algorithm is applied to establish a chiller width learning model to make a preliminary prediction of the chiller operation status; Step 2: Based on the actual operation data of the chiller and the law of conservation of energy, the mean square error of the cooling water outlet temperature of the width learning model, the mean square error of the chiller electric power, and the energy loss function of the chiller are calculated to establish a custom physical loss function for the chiller; Step 3: Use the error back propagation mechanism and Adam optimizer to optimize the chiller’s custom physical loss function, update the weights and bias coefficients of the feature layer and enhancement layer in the width learning algorithm, improve the accuracy of the width learning model, and establish a physically constrained width learning model. Step 4: Establish a priori knowledge model of the chiller; Step 5: Combine the chiller prior knowledge model with the physical constraint width learning model to establish a chiller extrapolation model based on physical constraint width learning and prior knowledge.

2. A modeling method for a chiller extrapolation model based on physical constraint width learning and prior knowledge as claimed in claim 1, characterized in that: The method further includes step 6, building a dynamic data-driven application system for online data collection and calculation, realizing real-time prediction of the chiller operation status and online update of the chiller extrapolation model; Specifically, the dynamic data-driven application system collects field sensor data and transmits it to the extrapolation model building module to achieve real-time prediction of the chiller's operating status; According to the error between the output result of the extrapolation model of the chiller and the actual measurement value, a custom physical loss function is calculated, and the modeling process is optimized using the error back propagation mechanism and Adam optimizer to achieve online correction of the accuracy of the extrapolation model.

3. A modeling method for a chiller extrapolation model based on physical constraint width learning and prior knowledge as claimed in claim 1, characterized in that: In step 1, the input variables of the width learning algorithm are the chilled water outlet temperature, the chilled water inlet temperature, the cooling water inlet temperature, the chilled water flow rate, and the cooling water flow rate, and the output variables of the width learning algorithm are the cooling water outlet temperature and the chiller power.

4. A modeling method for a chiller extrapolation model based on physical constraint width learning and prior knowledge as claimed in claim 1, characterized in that: In step 2, the mean square error of the cooling water outlet temperature reflects the error between the predicted value and the actual value of the cooling water outlet temperature of the chiller under the historical operating parameters of the chiller. The mean square error MSE of the cooling water outlet temperature is TCO It is expressed as: Where n is the number of samples; T cw,L is the actual value of cooling water outlet temperature, ℃; is the predicted value of the cooling water outlet temperature of the chiller model, ℃.

5. A modeling method for a chiller extrapolation model based on physical constraint width learning and prior knowledge as claimed in claim 1, characterized in that: In step 2, the mean square error of the chiller power reflects the error between the predicted value and the actual value of the chiller power under the historical operating parameters of the chiller. The mean square error MSE of the chiller power p It is expressed as: Where n is the number of samples; P is the actual value of the chiller power, kW; is the predicted value of the chiller's electrical power, kW.

6. A modeling method for a chiller extrapolation model based on physical constraint width learning and prior knowledge as claimed in claim 1, characterized in that: In step 2, the chiller energy loss function reflects the degree to which the model prediction value deviates from the physical constraint. energy It is expressed as: Where c is the specific heat capacity of chilled water, KJ / (kg·℃); G ew is the flow rate of chilled water, L / s; T ew,E is the chilled water inlet temperature, °C; T ew,L is the chilled water outlet temperature, °C; is the predicted value of the chiller power, kW; G cw is the cooling water flow rate, L / s; is the predicted value of cooling water outlet temperature, ℃; T cw,E is the cooling water inlet temperature, °C; min is the lower limit of the fluctuation range, kW; max is the upper limit of the fluctuation range, kW.

7. A modeling method for a chiller extrapolation model based on physical constraint width learning and prior knowledge as claimed in claim 1, characterized in that: In step 2, the chiller's custom physical loss function is expressed as: CustomLoss=MSE TC0 +MSE p +Loss energy Where CustomLoss is the physical loss of the chiller; MSE TC0 MSE is the mean square error of the cooling water outlet temperature; p is the mean square error of the chiller power; Loss energy is the energy loss of the chiller.

8. A modeling method for a chiller extrapolation model based on physical constraint width learning and prior knowledge as claimed in claim 1, characterized in that: In the step 3, the error back propagation mechanism and Adam optimizer are used to optimize the custom physical loss function of the width learning model to establish a physical constraint width learning model; the back propagation mechanism and Adam optimizer are used to update the feature layer weights and enhancement layer weights in the width learning network according to the custom physical loss function to minimize the custom physical loss function; the output layer weights are solved using a pseudo-inverse solution method.

9. A modeling method for a chiller extrapolation model based on physical constraint width learning and prior knowledge as claimed in claim 1, characterized in that: In step 4, the chiller prior knowledge model is expressed as: Q ew =c·G ew ·(T ew，E -T ew，L ) P=a0+a1·PLR+a2·PLR 2 +a3·PLR 3 COP=b0+b1·PLR+B2·PLR 2 +b3·PLR 3 +b4·PLR 4 In the formula, Q ew is the cooling capacity of the evaporator, kW; c is the specific heat capacity of chilled water, KJ / (kg·℃); G ew is the flow rate of chilled water, L / s; T ew，E is the return water temperature of chilled water, °C; T ew，L is the supply temperature of chilled water, °C; PLR is the partial load rate of the chiller; Q n is the rated cooling capacity of the chiller, kW; P is the compressor power, kW; COP is the energy efficiency ratio of the chiller; a0-a3, b0-b4 are the coefficients in the performance curve; T cw，L is the cooling water outlet temperature, ℃; G cw is the cooling water flow rate, L / s; T cw，E is the cooling water inlet temperature, ℃.

10. A modeling method for a chiller extrapolation model based on physical constraint width learning and prior knowledge as claimed in claim 1, characterized in that: In step 5, the prior knowledge model is integrated with the physical constraint width learning model, that is, the output of the prior knowledge model and the output of the physical constraint width learning model are weighted averaged, which is expressed as: Where Y KPCBLS is the output of the extrapolation model; Y K is the output of the prior knowledge model; Y PCBLS The output of the model is learned for the physical constraint width.

Images