Method for controlling water chiller and control system, equipment, storage medium and program product thereof

Abstract

The invention provides a method for controlling a water chiller in a water chilling unit and a control system, equipment, a storage medium and a program product thereof, is used for improving the energy efficiency of the water chilling unit, and relates to the technical field of energy management. The method for controlling the water chiller in the water chiller unit comprises the steps that the load of the water chiller is predicted based on the time characteristic, the weather characteristic and the historical cooling load characteristic of the water chiller; according to the load prediction result and the partial load rate constraint condition, a control sequence used for controlling the water chiller is determined, and the control sequence comprises the starting number and the operation sequence of the water chiller in the water chiller unit; selecting an optimal control sequence from the control sequences according to the cooling-water machine model; and controlling the operation of the cooling-water machine based on the optimal control sequence.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of energy management, in particular to the technical field of optimizing the operation of a water chiller, and especially to a method for controlling a water chiller in a water chiller unit, a control system, a device, a storage medium and a program product. BACKGROUND

[0002] HVAC system (Heating Ventilation and Air Conditioning) is a comprehensive environmental control system integrating heating, ventilation and air conditioning regulation. It is crucial to reduce the energy consumption of HVAC systems in large commercial buildings. The energy consumption of a water chiller unit accounts for more than 50% of the energy consumption of an HVAC system. The water chiller unit is responsible for providing cooling capacity to meet the cooling demand of the building space.

[0003] In large commercial buildings, a water chiller unit including multiple water chillers is usually used to meet the cooling load. The sequencing of multiple water chillers in a water chiller unit is a control strategy aimed at optimizing the on or off state of multiple water chillers. By strategically controlling the on or off state of the water chiller, the water chiller can operate at a high COP (Coefficient of Performance) to meet the cooling demand while achieving significant energy savings. SUMMARY

[0004] The embodiments of the present disclosure propose a method for controlling a water chiller in a water chiller unit, a control system, a device, a storage medium and a program product to improve the energy efficiency of the water chiller unit.

[0005] In a first aspect, the embodiments of the present disclosure propose a method for controlling a water chiller in a water chiller unit, comprising: predicting the load of the water chiller based on time characteristics, weather characteristics and historical cooling load characteristics of the water chiller; determining a control sequence for controlling the water chiller according to the load prediction result and a partial load rate constraint condition, the control sequence including the number of water chillers turned on and the operation sequence in the water chiller unit; selecting an optimal control sequence from the control sequence according to a water chiller model; and controlling the operation of the water chiller based on the optimal control sequence.

[0006] In some embodiments, predicting the load of the water chiller based on time characteristics, weather characteristics and historical cooling load characteristics of the water chiller comprises: receiving time characteristics, weather characteristics and historical cooling load characteristics of the water chiller through an XGBoost model; outputting a cooling load demand prediction curve through the XGBoost model; and predicting the load of the water chiller based on the prediction curve.

[0007] In some embodiments, the method further comprises establishing a water chiller model according to historical operation data of the water chiller.

[0008] In some embodiments, the chiller model is established according to historical operation data of the chiller, comprising: determining target feature parameters based on initial feature parameters related to the chiller; and in response to the artificial neural network receiving the target feature parameters, establishing the chiller model and outputting energy consumption and coefficient of performance of the chiller, wherein the initial feature parameters comprise at least one of chiller set operation list, chilled water flow, cooling water flow, chiller set temperature point data, outdoor ambient temperature, total cooling load and partial load ratio.

[0009] In some embodiments, the target feature parameters are determined based on the initial feature parameters, comprising: determining an effective range of the initial feature parameters by using a statistical method of covariance matrix and confidence ellipse, and determining the initial feature parameters within the effective range as the target feature parameters.

[0010] In some embodiments, the optimal control sequence is selected from the control sequences according to the chiller model, comprising: determining energy consumption of the chiller set corresponding to the control sequences within a specific time interval according to the chiller model; calculating energy consumption of each control process in any control sequence according to the chiller model; and selecting the optimal chiller control sequence based on a comparison result of energy consumption of the plurality of chiller sets and energy consumption of each control step in the control sequence for controlling the chiller set to work.

[0011] In some embodiments, before the energy consumption of each control process in any control sequence is calculated, the method further comprises: establishing a state transition equation to determine the control sequence and cumulative energy consumption of the control sequence from a start time to a time t, a control action in the control sequence determining an on or off state of the chiller at the next time, wherein t is a positive integer; and storing a tuple formed by the control sequence and the cumulative energy consumption if the tuple is not stored.

[0012] In some embodiments, the optimal control sequence is selected from the control sequences according to the chiller model, comprising: searching for a control sequence satisfying the cooling load and operation constraint conditions when the partial load ratio meets a first condition; and searching for a control sequence satisfying the cooling load and operation constraint conditions when the partial load ratio meets a second condition if no control sequence is searched. Wherein the numerical range of the second condition is greater than the numerical range of the first condition, and the numerical range of the first condition is a range in which the partial load ratio is close to the maximum value of the coefficient of performance.

[0013] In some embodiments, the control sequence for controlling the chiller is determined according to the load prediction result and the partial load ratio constraint condition, comprising: generating a plurality of initial control sequences according to the cooling load demand prediction curve, the initial control sequences being limited by minimum on and off intervals; and determining the control sequence from the plurality of initial control sequences based on the partial load ratio constraint condition.

[0014] In a second aspect, the embodiments of the present disclosure provide a control system for controlling a chiller in a chiller unit, comprising: a load prediction module, a screening module, an optimization module and a control module. The load prediction module is configured to predict a load of the chiller based on time characteristics, weather characteristics and historical cooling load characteristics of the chiller. The screening module is configured to determine a control sequence for controlling the chiller according to the load prediction result and a partial load rate constraint condition, the control sequence comprising a number of chiller starts and an operation sequence of the chiller in the chiller unit. The optimization module is configured to select an optimal control sequence from the control sequence according to a chiller model. The control module is configured to control the chiller to operate based on the optimal control sequence.

[0015] In some embodiments, the load prediction module is specifically configured to receive the time characteristics, the weather characteristics and the historical cooling load characteristics of the chiller through an XGBoost model, output a cooling load demand prediction curve through the XGBoost model, and predict the load of the chiller based on the prediction curve.

[0016] In some embodiments, the optimization module further comprises a chiller model trainer configured to determine target feature parameters based on initial feature parameters related to the chiller, establish a chiller model in response to the artificial neural network receiving the target feature parameters, and output energy consumption and a coefficient of performance of the chiller. The initial feature parameters include at least one of a chiller operation table, a chilled water flow, a cooling water flow, chiller unit temperature point data, an outdoor environment temperature, a total cooling load and a partial load rate.

[0017] In some embodiments, the chiller model trainer is specifically configured to determine an effective range of the initial feature parameters using a statistical method of a covariance matrix and a confidence ellipse, and determine the initial feature parameters within the effective range as the target feature parameters.

[0018] In some embodiments, the optimization module is specifically configured to determine energy consumption of the chiller unit corresponding to the control sequence within a specific time interval according to the chiller model, calculate energy consumption of each control process in any control sequence according to the chiller model, and select an optimal chiller control sequence based on a comparison result of the energy consumption of the plurality of chiller units and the energy consumption of each control step in the control sequence for controlling the chiller unit to operate.

[0019] In some embodiments, the optimization module is further configured to establish a state transition equation to determine the control sequence and cumulative energy consumption of the control sequence from a start time to a time t, a control action in the control sequence determines an on or off state of the chiller at a next time, wherein t is a positive integer, and store a tuple formed by the control sequence and the cumulative energy consumption in a case where the tuple is not stored.

[0020] In some embodiments, the optimization module is further configured to search for the control sequence satisfying the cooling load and operation constraint condition when the partial load ratio meets a first condition, and search for the control sequence satisfying the cooling load and operation constraint condition when the partial load ratio meets a second condition if the control sequence is not searched. The numerical range of the second condition is greater than the numerical range of the first condition, and the numerical range of the first condition is a range in which the partial load ratio is close to the maximum value of the performance coefficient.

[0021] In some embodiments, the screening module is specifically configured to generate a plurality of initial control sequences according to the cooling load demand prediction curve, the initial control sequences being limited by the minimum opening and closing interval, and determine the control sequence from the plurality of initial control sequences based on the partial load ratio constraint condition.

[0022] In a third aspect, an electronic device is provided, which includes at least one processor, and a memory connected with the at least one processor in communication; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to implement the water chiller control method described in any implementation manner of the first aspect when executed.

[0023] In a fourth aspect, a non-transitory computer-readable storage medium storing computer instructions is provided, and the computer instructions are used to enable a computer to implement the water chiller control method described in any implementation manner of the first aspect when executed.

[0024] In a fifth aspect, a computer program product including a computer program is provided, and the computer program is used to enable a processor to implement the water chiller control method described in any implementation manner of the first aspect when executed.

[0025] The method for controlling the water chiller in the water chiller unit provided by the present disclosure can predict the load of the water chiller based on the time characteristics, weather characteristics and historical cooling load characteristics of the water chiller, and can only use commonly used water chiller sensor points such as the temperature and flow rate of the chilled water and cooling water and the power of the water chiller, without the need for additional installation of measuring equipment, thereby ensuring that the present disclosure can be deployed on a large scale. In addition, the optimal control sequence is selected from the control sequence according to the water chiller model, which can consider the dynamic influence of each step of the optimization strategy, thereby ensuring that the optimal control sequence is accurately obtained, and the optimal control sequence can improve the energy efficiency of the water chiller unit while meeting the refrigeration demand.

[0026] It should be understood that the content described in this part is not intended to identify key or important features of the embodiments of the present disclosure, nor is it used to limit the scope of the present disclosure. Other features of the present disclosure will become apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0027] Other features, objects, and advantages of the present disclosure will become more apparent from the following detailed description of non-limiting embodiments made with reference to the accompanying drawings: Figure 1 A flow chart of a method for controlling a chiller in a chiller unit according to an embodiment of the present disclosure; Figure 2 A detailed work flow chart of a method for controlling a chiller in a chiller unit according to an embodiment of the present disclosure; Figure 3 A schematic diagram of a chiller model according to an embodiment of the present disclosure; Figure 4 A schematic diagram of a confidence ellipse according to an embodiment of the present disclosure; Figure 5 A schematic diagram of a two-step optimization method according to an embodiment of the present disclosure; Figure 6a A flow chart of a two-step optimization method according to an embodiment of the present disclosure; Figure 6b A flow chart of a dynamic programming method according to an embodiment of the present disclosure; Figure 7 A schematic diagram of a dynamic programming according to an embodiment of the present disclosure; Figure 8 A schematic diagram of a control module of a method for controlling a chiller in a chiller unit according to an embodiment of the present disclosure; Figure 9 A schematic diagram of simulation results of chiller sequencing optimization according to an embodiment of the present disclosure; Figure 10 A schematic diagram of average duration of computation using an ANN model and a heuristic model, respectively; Figure 11a and Figure 11b A schematic diagram of comparison of number of feasible control sequences before and after optimization using a two-step optimization method, respectively; Figure 12a and Figure 12b A schematic diagram of comparison of computation time before and after using a DP method, respectively; Figure 13 A schematic diagram of computation time corresponding to chiller sequencing optimization using a DP method, a combination of a DP method and a two-step optimization, and an exhaustive search method, respectively; Figure 14 A schematic diagram of average and extreme values of computation time corresponding to chiller sequencing optimization using a DP method, a combination of a DP method and a two-step optimization, and an exhaustive search method, respectively. DETAILED DESCRIPTION

[0028] Exemplary embodiments of the present disclosure are described below with reference to the accompanying drawings. Although various detailed descriptions are provided in these embodiments for the ease of understanding, it will be understood by one of ordinary skill in the art that these embodiments and their detailed descriptions are merely exemplary, and various modifications can be made to these embodiments without departing from the teachings of the present disclosure. Also, detailed descriptions of well-known functions and structures incorporated herein will be omitted for the sake of clarity and conciseness. Furthermore, the embodiments in the present disclosure and features in the embodiments can be combined with each other if there is no conflict.

[0029] Traditional methods of controlling chiller in a chiller plant, such as Rule-Based-Control (RBC), are often not the best control method. RBC usually regulates the on-line and off-line status of the chiller when the measured cooling load exceeds a pre-set threshold or according to the experience of the operator. This method has two problems, the first problem is that it cannot use real-time operation data or prediction information such as weather forecast to optimize the control strategy in time. The second problem is that the COP of the chiller and the actual cooling capacity are not constant, but will fluctuate with the change of the operating conditions. For example, centrifugal chillers are more efficient when close to full load, while screw chillers usually reach the highest efficiency at part load. However, the traditional RBC method usually assumes that the cooling capacity of the chiller is equal to its rated capacity when regulating the chiller. This idealized assumption is likely to cause the chiller to be unable to run in the optimal state in the complex and variable actual operating environment, thereby causing the overall efficiency to be reduced.

[0030] Model Predictive Control (MPC) has become the state-of-the-art solution to optimize the control sequence of chiller plants. By integrating system models to predict future cooling loads, MPC provides a dynamic control strategy that maximizes energy savings while ensuring that operational constraints are met. In current approaches, the constructed models can include a load prediction model and a chiller model. The load prediction model predicts the chiller plant load for a future time period based on current building operation conditions and weather forecast information. The chiller model describes the chiller performance, reflecting the relationship between chiller operating conditions and chiller COP, which is defined as the ratio of chiller cooling capacity to its input power. In practice, the COP of a chiller often fluctuates with changes in operating conditions, such as the Partial Load Ratio (PLR). The MPC controller will preferentially select the chiller with the highest COP in the chiller plant to ensure efficient operation while meeting the predicted cooling requirements. Therefore, the accuracy of the load prediction model and the chiller model directly affects the optimization performance of MPC, and common models include physical models, grey box models, and black box models.

[0031] Physical models are based on thermodynamics and heat transfer theory, constructed by collecting technical specifications and physical details of building or chiller equipment components. The constructed model can accurately describe the operating characteristics of the chiller. Grey box models are an improvement over physical models, relying mainly on operational data and using statistical techniques to determine parameters.

[0032] However, physical models and grey box models are less commonly used in practice due to the following problems. The first problem is that physical models and grey box models require a complex system of parameters to construct and operate, and in large commercial buildings, collecting these parameters poses a significant challenge. Due to the lack of measurement points in central cooling systems, it is difficult to obtain the required parameter information comprehensively and accurately. The second problem is that physical models and grey box models require a large amount of professional knowledge to establish the model. The third problem is that physical models and grey box models do not take into account the dynamic changes of the building environment and chiller plant. They only optimize the chiller control sequence for a single time step, without considering the impact of optimization on subsequent time step load changes and chiller performance. Therefore, physical models and grey box models often struggle to find the optimal chiller control sequence.

[0033] With building management systems becoming more and more advanced, it is possible to collect operational data by the hour or even by the minute. Black-box models (usually data-driven models) have attracted attention. Black-box models do not require explicit physical modeling, but learn building dynamics from measured data, and have high flexibility. However, current black-box models have the following two problems. The first problem is that there is uncertainty in the data fitting model, and the control performance is only reliable within the trust region. Beyond this range, the accuracy will drop sharply, that is, there is a problem of data extrapolation. The second problem is that the model has high computational requirements, which may hinder real-time deployment. For example, the time required for calculation may exceed the control interval.

[0034] Another aspect of MPC controller design is the choice of optimization method. Common optimization methods for control sequences of chillers can be roughly divided into heuristic algorithms and global optimization algorithms. Heuristic algorithms usually simulate natural evolution processes to approach optimal solutions. Current heuristic algorithms, although computationally efficient, are very suitable for large-scale problems, but often only find suboptimal chiller control sequences. Global optimization algorithms, such as linear programming (LP) and mixed-integer linear programming (MILP), can solve the problem of chiller control sequences, but have the problem of low computational efficiency, especially when the number of chillers and the control range increases, which will result in too long optimization time and cannot be practically applied. In addition, the increase in model complexity also greatly exacerbates this problem, limiting the real-time deployment of MPC methods.

[0035] Based on the problems in the current method of controlling chillers in a chiller unit, the present disclosure provides a new method of controlling chillers in a chiller unit to accurately and quickly find the optimal control sequence in the control sequence of the chiller, so as to improve the energy efficiency of the chiller unit.

[0036] Embodiments of the present disclosure provide a method of controlling a chiller in a chiller unit, as shown in Figure 1 The method comprises: S101, predicting the load of the chiller based on the time characteristics, weather characteristics and historical cooling load characteristics of the chiller.

[0037] S102, determining a control sequence for controlling the chiller according to the load prediction result and the partial load rate constraint condition, the control sequence comprising the number of chiller units in the chiller unit and the operation sequence.

[0038] S103, selecting an optimal control sequence from the control sequence according to the chiller model.

[0039] S104, controlling the operation of the chiller based on the optimal control sequence.

[0040] The method for controlling the chiller in the chiller plant provided by the above-mentioned embodiments can predict the load of the chiller based on the time characteristics, weather characteristics and historical cooling load characteristics of the chiller, and can only use commonly used chiller sensor points such as the temperatures and flow rates of chilled water and cooling water and the power of the chiller, without the need for additional installation of measuring equipment, thereby ensuring that the present disclosure can be deployed on a large scale. In addition, the optimal control sequence in the control sequence is selected according to the chiller model, the dynamic influence of each step of the optimization strategy can be considered, and the optimal control sequence can be accurately obtained, and the optimal control sequence can be obtained while meeting the refrigeration demand and improving the energy efficiency of the chiller plant.

[0041] The chiller sequencing optimization problem aims to determine the optimal control sequence in the control sequence for controlling the chiller, which can maximize the total energy consumption while meeting the refrigeration demand and operating constraints. The objective function is to minimize the total power consumption of all chillers in the control range. This requires accurate prediction of the cooling load of the chiller and determination of the optimal operation sequence of the chiller in each control time step (i.e., determination of the optimal control sequence of the chiller in the chiller plant in each control step), so as to meet the cooling load with the minimum energy consumption. The control step refers to the interval of executing the control instruction. The optimization objective is shown in formula (1.a):

[0042] wherein E t is the total energy consumption of the chiller at time t, which is obtained from the chiller model, and N is the number of time intervals. To ensure the feasibility and practicality of the chiller sequencing optimization problem, the optimization problem needs to meet the constraints in formulas (1.b)-(1.e).

[0043]

[0044] wherein CL t is the cooling load of the chiller plant, PLR t is the partial load ratio of the i-th chiller at time t, is the capacity of the i-th chiller, M is the number of chillers, is the predicted power of the chiller model, x t is the input value of the chiller model at time t, and w is the parameter of the chiller model. In addition, the ratio of MUT (Minimum Up Time) and MDT (Minimum Down Time) reflects the physical limitations of the chiller, and the ratio needs to meet certain conditions to prevent frequent start-up and shutdown of the chiller. In addition, the PLR of each chiller must be kept within a specified range to maintain the high efficiency of the chiller.

[0045] Control action vector u t is the control sequence of the multiple chillers at time t, which determines the on or off state of the chillers at the next time, so s t+1 = u t , where s t+1 is the state vector of the multiple chillers at time t+1. Assuming that the PLR of all chillers in the on state is the same, the total refrigeration capacity provided by the chillers at any given time must meet the predicted cooling load, which can ensure that the cooling load of the building is always met.

[0046] As Figure 2 shown, the key of the chiller sequencing optimization problem lies in quickly and accurately finding the optimal control sequence in the control sequences for controlling the chillers. The purpose of the screening process is to screen the feasible control sequences from all control sequences, that is, to determine the control sequences for controlling the chillers in S102 described above. The purpose of the optimization process is to quickly and accurately select the optimal control sequence from the feasible control sequences, that is, to select the optimal control sequence in S103 described above. Figure 2 The chiller model in Figure 3 may be referred to as shown.

[0047] In some embodiments, the prediction of the load of the chiller based on the time feature, the weather feature and the historical cooling load feature of the chiller in S101 described above comprises: receiving the time feature, the weather feature and the historical cooling load feature of the chiller by the XGBoost model; outputting a cooling load demand prediction curve by the XGBoost model, and predicting the load of the chiller based on the prediction curve, as shown in Figure 2 .

[0048] In the chiller ranking optimization problem, accurate prediction of chiller load is crucial because accurate load prediction can provide insights into future cooling demand and data support and decision-making basis for the efficient operation of chillers. The load prediction model provides important input data for the subsequent optimization process. This disclosure selects the Extreme Gradient Boosting (XGBoost) method to construct the load prediction model. The XGBoost model is an ensemble algorithm based on decision trees. Due to its high accuracy and efficient training speed, it is widely used in various prediction tasks, such as predicting financial electricity load and building energy efficiency. The XGBoost model can effectively handle large-scale data and high-dimensional features in building load prediction tasks and provide reliable predictions. The XGBoost model effectively handles large-scale data and high-dimensional features in load prediction tasks by continuously combining multiple predictors, thereby providing reliable predictions. Specifically, the new predictor improves prediction accuracy by focusing on and correcting the prediction error of the previous predictor. To prevent overfitting, the XGBoost model integrates multiple weak predictors (usually shallow decision trees) and regularization terms in the objective. The goal of the XGBoost model is to minimize the objective function of each generation, as shown in Equation (2):

[0049] Where i is the i-th sample to be predicted, n is the total number of samples, and t is the t-th iteration. It is a real label and prediction labels The loss function between It is the base learner added in the t-th iteration. It is a feature of the i-th sample. It is a regularization term used to avoid overfitting. Let represent the objective function at the t-th iteration.

[0050] The XGBoost model describes the relationship between input data and the future cooling load of a building, making the selection of input data crucial. A building's cooling load is caused by internal and external heat gain. Internal heat gain is primarily caused by occupant behavior, appliance usage, and lighting, and typically depends on the building's operating schedule. For example, typical commercial building usage is usually from 10 AM to 10 PM, and appliance usage often increases on weekends and holidays. External heat gain generally refers to solar radiation, air convection heat transfer, and heat conduction, and is mainly affected by outdoor weather conditions. Furthermore, due to the building's thermal mass, the current cooling load is influenced by the cooling load of previous periods; therefore, the input features of the XGBoost model can include time features, weather features, and historical cooling load features of the chiller.

[0051] For example, the cooling load exhibits a clear time cycle, and the hour, day of week, and month features can provide information on intra-day variation, weekday versus weekend differences, and seasonal variation, so the time features can be selected as the hour, day of week, and month features.

[0052] For example, the weather directly affects the cooling load, and the temperature, humidity, and solar radiation features help the XGBoost model capture the variation of the cooling load, so the weather features can be selected as the temperature, humidity, and solar radiation features.

[0053] For example, the cooling load has a strong time dependence, and inputting historical cooling load data can improve the accuracy and reliability of the XGBoost model.

[0054] Table 1 shows the input features of the XGBoost model. Among the time-related time features, the hour, day of week, and month can include data transformed by sine (sin) and cosine (cos) in addition to the original data. In addition, the wind direction, azimuth, and elevation angle in the weather features can also include data transformed by sine (sin) and cosine (cos) in addition to the original data. In addition to using the original data, some feature data in the time-related time features and weather features can use sine-transformed data and cosine-transformed data to better capture the periodic variation of the cooling load.

[0055] Table 1

[0056] All features in Table 1 can be normalized using the data standardization method (StandardScale) r. The Coefficient of Variation of Root Mean Square Error (CVRMSE) is used to evaluate the accuracy of the XGBoost model, as defined in Equations (3.a)-(3.b). The Root Mean Squared Error (RMSE) measures the average of the prediction error, while the CVRMSE normalizes the RMSE to make it a relative measure of prediction accuracy. The CVRMSE is used to evaluate the XGBoost model because it provides a standardized index that allows for more meaningful comparisons between different data sets.

[0057]

[0058] where y is the true cooling load value, is the predicted cooling load value. n is the number of samples. is the average cooling load value.

[0059] Hyperparameter tuning is essential to improve the performance of the XGBoost model. Bayesian Optimization (BO) as the most advanced hyper-parameter optimization (HPO) algorithm has higher efficiency and convergence. Specifically, the disclosure selects the Tree-Structured Parzen Estimator (TPE) in the Hyperopt library as the HPO algorithm. When using HPO to optimize the hyperparameters of the XGBoost model, a reasonable search space needs to be defined, including the maximum depth, learning rate, estimator number and regularization parameter and other hyperparameters, and Table 2 shows the range of hyperparameter search.

[0060] Table 2

[0061] In some embodiments, the method for controlling the chiller in the chiller unit of the disclosure further comprises: establishing a chiller model according to historical running data of the chiller. For example, the historical running data of the chiller is updated once a week. The chiller model is used to describe the relationship between the operating conditions of the chiller unit and its COP.

[0062] In some embodiments, the chiller model is established according to the historical running data of the chiller, comprising: determining a target feature parameter based on an initial feature parameter related to the chiller; and in response to the artificial neural network receiving the target feature parameter, establishing the chiller model and outputting the energy consumption and the coefficient of performance of the chiller. Wherein, the initial feature parameter includes at least one of the chiller unit operation list, the chilled water flow, the cooling water flow, the chiller unit temperature point data, the outdoor environment temperature, the total cooling load and the partial load rate.

[0063] In some embodiments, based on the initial feature parameter, the target feature parameter is determined, comprising: using the statistical method of covariance matrix and confidence ellipse to determine the effective range of the initial feature parameter, and determining the initial feature parameter within the effective range as the target feature parameter. The chiller model in the disclosure is constructed based on the target feature parameter (i.e. the artificial neural network method combined with the trust region is used in the construction of the chiller model of the disclosure), so that data interpolation can be avoided and the accuracy of the chiller model is improved.

[0064] The chiller model in the disclosure is described in detail below.

[0065] For the chiller, COP represents the ratio of cooling capacity to the electrical energy input required for the chiller to produce cooling, as shown in equation (4).

[0066]

[0067] Artificial Neural Network (ANN) is good at handling multivariate and nonlinear problems, and can automatically extract features from a large amount of data for dynamic chiller energy consumption prediction. The objective function J(w) of the ANN model is to continuously update the weight parameter w, thereby narrowing the gap between the predicted value and the true value, as shown in equation (5).

[0068]

[0069] where P is the sample size, x is the input of the ANN model, is the predicted value, is the true value.

[0070] The input of the ANN model (i.e., the initial feature parameters) can include at least one of the chiller set running list (s), the chilled water flow rate , the cooling water flow rate , the chilled water supply temperature , the cooling water return temperature , the outdoor environment temperature , the total cooling load and the partial load ratio (PLR). These parameters directly affect the energy efficiency and cooling load variation of the chiller. Table 3 shows the explanation of the input parameters of the chiller model. By inputting these parameters, the ANN model can effectively predict the energy consumption and COP of the chiller during operation, thereby optimizing the overall energy consumption of the chiller unit, Figure 3 is a schematic diagram of the chiller ANN model. In addition, the hyperparameters of the ANN model are shown in Table 4.

[0071] Table 3

[0072] Table 4

[0073] However, data-driven models (such as ANN models) often encounter data extrapolation problems. When the input data exceeds the range of the training data, the performance of the ANN model becomes very unreliable. In order to increase the reliability of the ANN model, the initial feature parameters are processed using the statistical method of covariance matrix and confidence ellipse to determine the effective range of the initial feature parameters, and the initial feature parameters within the effective range are determined as the target feature parameters. The target feature parameters are combined with the ANN model, i.e., the ANN method combined with the trust region is proposed to increase the reliability of the ANN model, which can avoid the data extrapolation problem and enhance the stability and reliability of the system.

[0074] The covariance matrix is used to describe the linear relationship between multiple variables. For the two variables of chilled water temperature difference (x) and chilled water temperature difference (y), the definition of the covariance matrix C(x, y) is shown in equation (6).

[0075]

[0076] Where Var(x) and Var(y) are the variances of x and y, respectively, and Cov(x, y) is the covariance of x and y, as shown in equation (7).

[0077]

[0078] According to the covariance matrix, the Pearson correlation coefficient p is calculated to control the rotation angle of the confidence ellipse, as shown in equation (8).

[0079]

[0080] The major and minor axes of the confidence ellipse are determined by the standard deviations of the variables, respectively:

[0081] Where, and are the eigenvalues of the covariance matrix. x is the confidence level, and is the multiple of the standard deviation (e.g. ).

[0082] After completing the affine transformation of the ellipse, multiple confidence ellipses are drawn according to different confidence intervals. The ellipse represents that 95% of the data is within the ellipse, which is the confidence interval of the variable, Figure 4 is a schematic diagram of the confidence ellipse, where The dashed line represents the confidence range of the chilled water temperature. When the variables of the feasible control sequence exceed this range, the calculation result of the operating condition will be considered unreliable.

[0083] In some embodiments, the present disclosure can also construct a water chiller model through a heuristic model, which is a second-order regression model based on empirical parameters. The quadratic term is introduced to better capture the nonlinear relationship between the COP of the water chiller, the load rate, the chilled water temperature, and the cooling water temperature, while ensuring the simplicity of the empirical model and reducing the computational complexity. The formula of the heuristic model is shown in equation (10):

[0084] Where PLR is the partial load rate of the water chiller, T ccw,s is the cooling water supply temperature, reflecting the heat dissipation capacity of the water chiller, T chw,s is the chilled water supply temperature, reflecting the cooling capacity of the water chiller. T ccw,s and Tchw,s The difference between them directly reflects the efficiency of the chiller, so these two variables are introduced into the COP variation model. The coefficients is the regression fitting parameter, calibrated by the least square method according to the operating data of the chiller unit, used to describe the comprehensive influence of the PLR and the cooling / freezing water temperature difference on the COP. Specifically, is the intercept term, is the load first-order term coefficient, quantifying the linear influence of the PLR on the COP, is the load second-order term coefficient, capturing the nonlinear influence of the PLR on the COP, is the temperature difference first-order term coefficient, is the temperature difference second-order term coefficient, describing the nonlinear influence of the ΔT on the COP, is the interaction term coefficient, describing the coupling effect of the PLR and the ΔT.

[0085] The initial chiller sequencing optimization method adopts the RBC method, which is the benchmark for comparison with the MPC method. The objective of the chiller sequencing optimization is to minimize the cumulative energy consumption, so the performance index is the energy saving rate, as shown in equation (11).

[0086]

[0087] where J RBC is the total energy consumption of the chiller sequencing optimization using the RBC method, J MPC is the total energy consumption of the chiller sequencing optimization using the MPC method.

[0088] In some embodiments, the control sequence for controlling the chiller is determined according to the load prediction result and the partial load rate constraint condition, including: generating a plurality of initial control sequences according to the cooling load demand prediction curve, the initial control sequences being limited by the minimum on and off interval; and determining the control sequence from the plurality of initial control sequences based on the partial load rate constraint condition.

[0089] For example, as shown in FIG. 3, Figure 2 all the control sequences in the set Figure 2 are initial control sequences, Figure 2 and the feasible control sequences in the set are the determined control sequences. The plurality of initial control sequences from t+1 to t+k can be generated according to the cooling load demand prediction curve, and the initial control sequences satisfying the partial load rate constraint condition are the determined feasible control sequences.

[0090] The present disclosure can select the optimal control sequence from a plurality of feasible control sequences by using a dynamic programming method and a two-step optimization method. The dynamic programming method can solve the problem of low computational efficiency caused by a large number of feasible control sequences. The two-step optimization method can solve the problem that the part load ratio constraint condition may be very strict, and no feasible control sequence can be found, thereby further optimization cannot be performed.

[0091] Each feasible control sequence F in the present disclosure is a control sequence with a length of N, as shown in equation (12).

[0092]

[0093] wherein s t is the combination of the switching state of the water chiller at time t, s i,t is the state of the i-th water chiller at time t, s i,t may be 0 or 1. N is the number of time intervals, which is calculated by equation (12.c), M is the total number of water chillers, T is the control range, and Δt is the control interval.

[0094] To solve the energy consumption of the feasible control sequence, the ANN model of each water chiller needs to be used at each time. Assuming that the number of feasible control sequences is Q, the total calculation amount is M N Q. In addition, the number of feasible control sequences increases with the increase of the time interval, resulting in an exponential increase in the calculation amount. Therefore, as the number of time intervals increases, the calculation time may become too long, which may exceed the length of the control time interval. In order to solve the problem of long calculation time, the present disclosure can use a dynamic programming method to find the optimal control sequence, which will be described in detail below.

[0095] In addition, when the part load ratio constraint condition is relatively strict, no feasible control sequence may be screened, for example, in order to ensure the COP of the water chiller set and limit the number of feasible solutions, the constraint range of the PLR is relatively narrow, and the case of no feasible control sequence is easy to occur. In order to solve this problem, the present disclosure can use a two-step optimization method to screen the feasible control sequence, which will be described in detail below.

[0096] In some embodiments, according to the water chiller model, the optimal control sequence is selected from the control sequence, comprising: searching for a control sequence satisfying the cooling load and operation constraint condition when the part load ratio satisfies a first condition; and when no control sequence is searched, searching for a control sequence satisfying the cooling load and operation constraint condition when the part load ratio satisfies a second condition. Wherein the numerical range of the second condition (such as the second range in Figure 5 ) is greater than the numerical range of the first condition (such as the first range in Figure 5The first condition is that the PLR is close to the maximum COP, and the numerical range of the first condition is the range where the PLR is close to the maximum COP. The two-step search method can improve the optimization efficiency and avoid the situation where there is no feasible control sequence.

[0097] Specifically, Figure 5 The two-step optimization process is shown. In the first step, the optimization process focuses on narrowing the range of PLR. By limiting the PLR, the configuration when the chiller is running is given priority to be closer to its optimal efficiency point (i.e., higher COP), i.e., searching when the PLR satisfies the first condition. In the second step, if the initial search under the PLR constraint condition does not find a control sequence that satisfies the cooling load and operation constraints, the constraint condition of PLR is relaxed to conduct a more extensive search, i.e., searching when the PLR satisfies the second condition. The specific process of two-step optimization is shown in Figure 6a As shown, the specific method of two-step optimization includes S201 to S204. If the control sequence obtained by the first search is empty, S203 is executed for the second search. If the control sequence obtained by the first search is not empty, S204 is directly executed.

[0098] The first step of the two-step search optimization method in the above embodiment greatly reduces the number of feasible control sequences needed, thereby speeding up the optimization process. The limited range of PLR is shown in formula (1.d). The second step of the two-step search optimization method relaxes the constraint condition of PLR to conduct a more extensive search, ensures that a feasible solution is found for subsequent optimization by considering more chiller running states, and the specific method is to replace formula (1.d) with formula (13).

[0099]

[0100] Dynamic programming (DP) is a method for solving complex optimization problems. It divides the original problem into overlapping subproblems and stores the solutions to these subproblems to avoid redundant calculations. DP is particularly suitable for multi-stage, sequential decision-making problems, especially when the objective function cannot be divided. Problems that can be effectively solved by DP usually have the following characteristics: first, optimal substructure, the optimal solution of a problem can be constructed from the optimal solutions of its subproblems. Second, subproblem overlap, the problem is repeatedly solved when recursively decomposed. Third, no aftereffect, once a subproblem is solved, it will not be affected by future decisions. The chiller sequencing optimization problem has these characteristics, so it is suitable for using the DP method.

[0101] In some embodiments, selecting the optimal control sequence from the control sequences according to the chiller model comprises: determining the energy consumption of the chiller units corresponding to the control sequences in a specific time interval according to the chiller model; calculating the energy consumption of each control step in any control sequence according to the chiller model; and selecting the optimal chiller control sequence based on the comparison of the energy consumption of the chiller units and the energy consumption of each control step in the control sequence that controls the operation of the chiller units.

[0102] In some embodiments, before calculating the energy consumption of each control step in any control sequence, further comprising: establishing a state transition equation to determine the control sequence and the cumulative energy consumption of the control sequence from the start time to time t, where the control action in the control sequence determines the on or off state of the chiller at the next time, and t is a positive integer; and storing the tuple formed by the control sequence and the cumulative energy consumption if it is not stored.

[0103] Specifically, selecting the optimal control sequence from the control sequences according to the chiller model comprises the following steps: First, decomposition of the original problem: the original problem is decomposed into several stages, each stage corresponds to a plurality of sub-problems, and the characteristics of the sub-problems are defined as states. In the chiller scheduling problem, the sub-problems involve determining the energy consumption of the chiller units in a specific time interval Δt, and the state is the combination of the on or off state of the chiller units at time t, and the decision variable is the on or off operation of each chiller unit.

[0104] Second, establishing a recursive relationship: this includes establishing a state transition equation. The optimal strategy has a feature that no matter the initial state and the initial decision, the remaining decisions must constitute an optimal strategy related to the state generated by the first decision. Therefore, the recursive relationship can be determined as:

[0105] where u t is the decision operation at time t, i.e. the state switching of the chiller units, E(u t ) is the energy consumption of the time interval after the operation u t is calculated by the ANN model, and V t (u t ) is the cumulative energy consumption from the start to time t.

[0106] Third, solving the sub-problems in order: this includes using the ANN model to solve the energy consumption of each chiller state combination in each time interval. Since this process involves repeatedly solving sub-problems, storing the solutions of previously calculated sub-problems can speed up the overall calculation. Figure 6b This shows how the dynamic programming method reduces the amount of calculation by storing and remembering sub-problems.

[0107] In some embodiments, the "solving sub-problems in sequence" includes: in each control time interval, solving the energy consumption for each chiller state combination, and storing the calculated sub-problem results to avoid repeated calculation and reduce the overall calculation amount; the process is as shown in Figure 6b , and corresponds to the node one-side structure in Figure 7 .

[0108] 1. Definition of state and transition

[0109] In some embodiments, the state is used to represent the time , the running combination of chillers, and the necessary constraint information, specifically including: the chiller switch vector ; the accumulated time related to the minimum up-time (MUT) and minimum down-time (MDT); the PLR determined by the running combination and the necessary characteristics (such as temperature, flow, etc.) for energy consumption estimation.

[0110] In some embodiments, the control action causes the state to transition from to , and only when the cooling load constraint, the PLR constraint, the MUT / MDT constraint, and the running constraint such as chiller on / off are met, the transition is considered feasible and is retained in the directed edge of Figure 7 .

[0111] 2. The first calculation step in Figure 6b (corresponding to the Figure 7 first layer)

[0112] In some embodiments, the "first calculation step" is not to calculate the total energy consumption of the entire sequence at once, but rather: generate the feasible state set at time ; ; for each , call the chiller model (such as the ANN model) to calculate the phase energy consumption of the time interval , and obtain the accumulated energy consumption initialization value .

[0113] Accordingly, the first calculation step in Figure 6b corresponds to the Figure 7 "phase energy consumption evaluation of the first layer node".

[0114] 3. The second calculation step in Figure 6b (cross-layer recursive merging)​

[0115] In some embodiments, the "second calculation step" is used to perform Bellman recursion: - for each feasible state at time , calculate

[0116] and record the predecessor state pointer prev that reaches the minimum value.

[0117] - sequentially advance to t+K in the above manner.

[0118] Accordingly, Figure 6b the second calculation step in corresponds to Figure 7 "feasible transition and cumulative value update across adjacent time instants" in and not the intra-time-instant comparison between parallel energy consumptions.

[0119] 4. The meaning of the (state) determination step and state structure

[0120] In some embodiments, Figure 6b the "determination state step" in is used to determine the feasibility of each state and its transition and to prune, including: the cooling load satisfies: the on-unit capacity (determined by PLR / unit type) is not less than the predicted load; the MUT / MDT satisfies: the on / off cumulative duration satisfies the minimum duration limit; the PLR interval: prefer to search within the first condition (PLR window close to the COP peak); if there is no solution, then relax according to the second condition; Figure 6a - and operational constraints such as upper limit of parallel units, minimum switching interval, etc.

[0121] When any constraint is not satisfied, the corresponding state or transition is directly discarded to reduce the subsequent calculation amount. The state structure relied on by the above determination is the aforementioned switching vector, timing information, and features required for energy consumption evaluation.

[0122] 5. Memoization storage and sub-problem reuse (the "storage and memory" of Figure 6b )

[0123] In some embodiments, to avoid repeated calls to the ANN model when recursively calculating subsequent layers, a cache entry is established for the calculated sub-problem and its result is stored . When the same appears again, the cache result is directly read without repeated solution, thereby significantly shortening the overall calculation time.​

[0124] 6. Termination and Search

[0125] In some implementations, at the predicted terminal time exist Select cumulative energy consumption The smallest, and along prev( The pointer is traced back in reverse to obtain the result from... to The optimal control sequence is obtained by employing the optimality principle of dynamic programming (optimal substructure overlaps with subproblems) and performing a comprehensive search of all feasible states and transitions (if necessary, according to...). Figure 6a (Relaxing PLR ensures feasibility), therefore the obtained control sequence is globally optimal.

[0126] By employing these steps and using the DP method, the chiller sequencing problem can be effectively optimized, minimizing energy consumption while ensuring that cooling requirements are met. Figure 7 This is a schematic diagram of the dynamic programming (DP) method. After finding the optimal control sequence for the first t+1 time steps, the subsequent feasible solutions will branch into different control sequences. Calculating the subproblem for the next time step requires recalculating the subproblems for all previous time steps. Therefore, the subproblems for previous time steps are stored to avoid redundant calculations and speed up the optimization process.

[0127] For example, exhaustive search can be chosen as a performance benchmark for dynamic programming (DP) methods. Exhaustive search finds the optimal solution by calculating the energy consumption of each feasible solution one by one. This method is chosen because it can find the global optimum, much like DP programming. However, this method requires calling an ANN model to evaluate each feasible solution, resulting in high computational costs.

[0128] For example, the methods described above in this disclosure can be encapsulated into a modular program for low-cost deployment and commercialization.

[0129] This disclosure also provides a control system for controlling the chiller in a chiller unit, such as... Figure 2 As shown, the system includes a load prediction module 101, a screening module 102, an optimization module 103, and a control module (not shown in the figure). The load prediction module 101 is configured to predict the chiller load based on time characteristics, weather characteristics, and the chiller's historical cooling load characteristics. The screening module 102 is configured to determine a control sequence for controlling the chillers based on the load prediction results and partial load factor constraints. The control sequence includes the number of chillers to be turned on and their operating order within the chiller unit. The optimization module 103 is configured to select the optimal control sequence from the control sequences based on the chiller model. The control module is configured to control the chiller operation based on the optimal control sequence.

[0130] In some embodiments, the load prediction module 101 is specifically configured to receive the time features, the weather features and the historical cooling load features of the water chiller through the XGBoost model, output a cooling load demand prediction curve through the XGBoost model, and predict the load of the water chiller based on the prediction curve.

[0131] In some embodiments, the optimization module 103 further comprises a water chiller model trainer configured to determine target feature parameters based on initial feature parameters related to the water chiller, establish a water chiller model in response to the artificial neural network receiving the target feature parameters, and output the energy consumption and the coefficient of performance of the water chiller, wherein the initial feature parameters comprise at least one of a water chiller operation list, a chilled water flow, a cooling water flow, water chiller set temperature point data, an outdoor environment temperature, a total cooling load and a partial load ratio.

[0132] In some embodiments, the water chiller model trainer is specifically configured to determine an effective range of the initial feature parameters by using a statistical method of a covariance matrix and a confidence ellipse, and determine the initial feature parameters within the effective range as the target feature parameters.

[0133] In some embodiments, the optimization module 103 is specifically configured to determine the energy consumption of the water chiller set corresponding to the control sequence in a specific time interval according to the water chiller model, calculate the energy consumption of each control process in any control sequence according to the water chiller model, and select the optimal water chiller control sequence based on a comparison result of the energy consumptions of the plurality of water chiller sets and the energy consumptions of each control step in the control sequence for operating the water chiller set.

[0134] In some embodiments, the optimization module 103 is further configured to establish a state transition equation to determine the control sequence and the cumulative energy consumption of the control sequence from a start time to a time t, wherein t is a positive integer, and store a tuple formed by the control sequence and the cumulative energy consumption in a case where the tuple is not stored, and a control action in the control sequence determines the on or off state of the water chiller at the next time.

[0135] In some embodiments, the optimization module 103 is further configured to search for a control sequence satisfying the cooling load and operation constraint conditions when the partial load ratio satisfies a first condition, and search for a control sequence satisfying the cooling load and operation constraint conditions when the partial load ratio satisfies a second condition in a case where the control sequence is not searched for; wherein the numerical range of the second condition is greater than the numerical range of the first condition, and the numerical range of the first condition is a range where the partial load ratio is close to the maximum value of the coefficient of performance.

[0136] In some embodiments, the screening module 102 is specifically configured to generate a plurality of initial control sequences according to the cooling load demand prediction curve, the initial control sequences being limited by the minimum on and off interval, and determine the control sequence from the plurality of initial control sequences based on the partial load rate constraint condition.

[0137] The optimization method of the water chiller provided by the present disclosure is further described below in combination with test results.

[0138] For example, the Ruihong Tiandi Solar Palace in Shanghai can be used as the test site, and the historical data can include building cooling load data, water chiller temperature point data, water chiller operation data, and corresponding meteorological data of the building in 2023, sampled every 15 minutes. The training set of the load prediction model and the water chiller model comes from the historical data from January 1, 2023 to October 15, 2023, excluding data from August 20 to August 27, and the test set only includes data from August 20 to August 27.

[0139] Figure 9 The simulation results are shown in the figure, which includes six subplots from top to bottom. The first subplot compares the cooling load values predicted by the load prediction model and the actual values observed by the tester. The CVRMSE of the load prediction model is 11.29%, indicating that the prediction accuracy of the load prediction model is relatively high. It should be noted that the predicted values are very close to the true values on weekdays, indicating that the load prediction model is very effective under normal operating conditions. However, the prediction accuracy may decrease on weekends, which may be due to changes in the operation of the mall on weekends.

[0140] Figure 9 The second subplot compares the actual power (i.e., the initial power), the ANN simulated power, and the optimized power (i.e., the optimal power). The actual power refers to the true power under the original control sequence, the ANN simulated power refers to the simulated power under the original control sequence, and the optimized power refers to the power under the optimal control sequence found by using the ANN model, the dynamic programming algorithm, and the two-step optimization method. The simulated value (red curve in the figure) and the actual value (purple curve in the figure) of the ANN simulated power obtained by using the ANN model are very close, indicating that the ANN model effectively simulates the characteristics of the water chiller. The optimized power (green curve in the figure) obtained by using the ANN model, the dynamic programming, and the two-step optimization algorithm is lower than the actual value (purple curve in the figure), indicating that the algorithm finds a control sequence with lower power, achieving effective energy saving. It should be noted that the ANN model may occasionally overestimate energy consumption, for example, on August 21.

[0141] Figure 9The third sub-plot in Figure 8 shows the variation of the daily average energy saving rate throughout the test period, with an average energy saving rate of 9.73% reflecting the effectiveness of the MPC in reducing energy consumption. While the energy saving rate is typically above 10% each day, the exception is Sunday, where the MPC consumes more energy than the original sequence. Table 5 summarizes the accuracy of the load forecasting model and the chiller model, as well as the energy saving effect achieved by the MPC method during the test period.

[0142] Table 5

[0143] Figure 9 The fourth sub-plot in Figure 8 shows the variation of the ambient temperature, which is relatively flat on Sundays, while the cooling load forecasting error is relatively large on Sundays, which may be related to the reduced energy saving effect.

[0144] It is worth noting that the load forecasting model has a significant impact on the control performance of the MPC. While the load forecasting model exhibits high accuracy on weekdays, its accuracy decreases on weekends, particularly on Fridays and Saturdays, which may be due to the significant changes in internal load patterns rather than external temperature. Possible solutions include, for example, increasing the size of the training data set and introducing more time-dependent or lagged features.

[0145] Figure 9 The fifth sub-plot in Figure 8 shows the average COP of the chiller control sequences before and after optimization, with the MPC algorithm always selecting a chiller control sequence with a higher COP than the original configuration, thereby optimizing energy efficiency. The sixth sub-plot shows a detailed example of the chiller control sequence before and after optimization on August 23. The MPC algorithm is able to adapt to different operating conditions and cooling loads, selecting the chiller control sequence with the highest COP at any given time. For example, at noon, chiller 2 and chiller 4 exhibit the highest COP under the current environmental conditions and cooling demand, while chiller 2 and chiller 3 are more efficient in the evening.

[0146] Figure 10 The time required to solve the same feasible solution using the ANN chiller model and the heuristic chiller model is compared. The computational demand of the ANN chiller model is 250 times that of the heuristic chiller model, so when the number of feasible solutions increases, the computational demand of the ANN chiller model will increase significantly.

[0147] Figure 11a and Figure 11bThe comparison of the number of feasible solutions (i.e., feasible control sequences) before and after using the two-step optimization method is shown. The results show that as the prediction range expands and the control interval decreases, the number of feasible control sequences increases rapidly. Specifically, before optimization, when the control time interval is set to 15 min and the prediction time interval is set to 4 h, the number of feasible control sequences is approximately 6.76e+7. However, after using the two-step optimization method, the number of feasible control sequences decreases to 3.51e+5, a reduction of 99.47%. This indicates that the two-step optimization method not only greatly reduces the number of feasible control sequences, but also speeds up the optimization process. The two-step optimization method first restricts the PLR to the vicinity of the COP peak, and then relaxes the range of PLR if no feasible control sequence is found, ensuring that a feasible control sequence can be found, thus greatly reducing the number of feasible control sequences and greatly reducing the computational burden of the subsequent DP method.

[0148] Figure 12a and Figure 12b The required calculation time before and after using the DP method is compared, assuming the same number of feasible control sequences. The DP method avoids repeated calculations of the ANN model by storing sub-problems, thereby improving computational efficiency. Therefore, even if the number of feasible control sequences remains the same or increases, the computational efficiency will improve. In the longest control range, the calculation time using the DP method is reduced by 99.98%, from nearly 8 days to only 138.24 seconds.

[0149] Figure 13 The calculation time corresponding to the DP method alone, the combination of the DP method and the two-step optimization, and the exhaustive search method when optimizing the chiller sequencing is shown. To ensure that the control performance is not affected, the calculation time of each time step of the MPC should not exceed a certain percentage of the control interval, which is set to 10% here, and the red dashed line in the figure represents the maximum acceptable calculation time per time step, which is 90 seconds. The calculation time of the DP method alone is mostly below 5 seconds, and the curve fluctuates less, indicating that this method performs well when handling load changes and is always within the acceptable range. The calculation time of the combination of the DP method and the two-step optimization is slightly lower than that of the DP method alone, indicating that the two-step optimization method can effectively reduce the calculation time. In contrast, the calculation time of the exhaustive search method is significantly higher and often exceeds the maximum acceptable time in most time steps.

[0150] Figure 14 The average and extreme values of the calculation time of the DP method alone, the combination of the DP method and the two-step optimization, and the exhaustive search method are shown. Figure 14The y-axis in FIG. 3 is a logarithmic scale. The average calculation time of the method combining the DP method with the two-step optimization is 3.61 seconds, and the maximum value is 6.81 seconds. The average calculation time of the DP method alone is close to that of the method combining the DP method with the two-step optimization, but the maximum value of the DP method alone is 12.29 seconds. The average calculation time of the exhaustive search method is close to 165.26 minutes, and the fluctuation is large, and the maximum value of the exhaustive search method is significantly higher, and the calculation efficiency in practical application is low. Overall, the DP method greatly speeds up the optimization process, and the two-step optimization improves the stability of the DP method and avoids the peak of the calculation time. The method combining the DP method with the two-step optimization has obvious advantages in calculation time and stability, and is suitable for practical control scenarios.

[0151] In summary, the method for controlling a chiller in a water chilling unit and the control system thereof provided by the present disclosure have the following beneficial effects: First, the present disclosure only uses commonly used chiller sensor points, such as the temperature and flow rate of chilled water and cooling water and the power of the chiller, without the need for additional installation of measurement equipment, ensuring that the present disclosure can be deployed on a large scale.

[0152] Second, the present disclosure uses the MPC method, which can consider the dynamic influence of each step of the optimization strategy to ensure that the optimal control sequence is obtained.

[0153] Third, the present disclosure uses the ANN method combined with the trust region, which avoids data interpolation and improves the accuracy of the chiller model.

[0154] Fourth, the present disclosure uses the method combining the two-step optimization and the dynamic programming, which greatly improves the calculation efficiency and ensures that the optimal control sequence can be quickly selected, thereby ensuring the feasibility of the practical deployment of the present disclosure.

[0155] According to an embodiment of the present disclosure, the present disclosure further provides an electronic device, which comprises at least one processor and a memory connected with the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to implement the chiller control method described in any of the above embodiments.

[0156] According to an embodiment of the present disclosure, the present disclosure further provides a readable storage medium storing computer instructions for enabling a computer to implement the chiller control method described in any of the above embodiments when the computer executes the computer instructions.

[0157] According to an embodiment of the present disclosure, the present disclosure also provides a computer program product, which, when executed by a processor, can implement the cold water machine control method described in any of the above embodiments.

[0158] It should be understood that the above steps can be reordered, or some additional steps can be added or some of the steps discussed above can be deleted according to the teachings of the embodiments of the present disclosure and actual needs. In addition, according to actual conditions, each step described in the present disclosure can be executed in parallel or sequentially, or in different orders, as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved, and the present disclosure is not limited herein.

[0159] The above detailed description does not constitute a limitation on the protection scope of the present disclosure. According to design requirements and other factors, various forms of modifications can be made to the above-described embodiments, including mutual or alternative between features. Any modification within the scope of the teachings of the present disclosure should be included in the protection scope of the present disclosure.

Claims

1. A method for controlling a chiller in a chiller unit, comprising: Based on time characteristics, weather characteristics, and historical cooling load characteristics of the chiller, the load of the chiller is predicted; Based on the load forecast results and partial load factor constraints, a control sequence for controlling the chillers is determined, the control sequence including the number of chillers to be turned on and their operating order in the chiller unit; Based on the chiller model, the optimal control sequence is selected from the control sequence; and The operation of the chiller is controlled based on the optimal control sequence.

2. The method according to claim 1, wherein, The method of predicting the load of the chiller based on time characteristics, weather characteristics, and historical cooling load characteristics of the chiller includes: The time characteristics, weather characteristics, and historical cooling load characteristics of the chiller are received through the XGBoost model; The XGBoost model outputs a cooling load demand prediction curve; and Based on the predicted curve, the load of the chiller is predicted.

3. The method according to claim 1, wherein, The method further includes: The chiller model is established based on the historical operating data of the chiller.

4. The method according to claim 3, wherein, The step of establishing the chiller model based on the historical operating data of the chiller includes: Based on the initial characteristic parameters associated with the chiller, the target characteristic parameters are determined; and In response to the artificial neural network receiving the target feature parameters, the chiller model is established, and the energy consumption and coefficient of performance of the chiller are output. The initial characteristic parameters include at least one of the following: chiller unit operation list, chilled water flow rate, cooling water flow rate, chiller unit temperature point data, outdoor ambient temperature, total cooling load, and partial load rate.

5. The control method according to claim 4, wherein, The step of determining the target feature parameters based on the initial feature parameters includes: The effective range of the initial feature parameters is determined by statistical methods using covariance matrix and confidence ellipse, and the initial feature parameters that fall within the effective range are determined as the target feature parameters.

6. The control method according to claim 1, wherein, The step of selecting the optimal control sequence from the control sequence based on the chiller model includes: Based on the chiller model, the energy consumption of the chiller unit corresponding to the control sequence is determined within a specific time interval; Based on the chiller model, calculate the energy consumption of each control process in any of the control sequences; and Based on the comparison results of the energy consumption of multiple chiller units and the energy consumption of each control step in the control sequence that controls the operation of the chiller unit, the optimal chiller control sequence is selected.

7. The control method according to claim 6, wherein, Before calculating the energy consumption of each control process in any of the control sequences, the method further includes: Establish state transition equations to determine the control sequence and the cumulative energy consumption of the control sequence from the start time to time t, where a control action in the control sequence determines the on or off state of the chiller at the next time step, and t is a positive integer; and If the tuple formed by the control sequence and the accumulated energy consumption is not stored, the tuple is stored.

8. The control method according to claim 1 or 6, wherein, The step of selecting the optimal control sequence from the control sequence based on the chiller model includes: When the partial load rate meets the first condition, search for a control sequence that satisfies the cooling load and operating constraints; and If the control sequence is not found, a control sequence that satisfies the cooling load and operating constraints is searched when the partial load rate meets the second condition. The numerical range of the second condition is greater than that of the first condition, and the numerical range of the first condition is the range where the partial load rate is close to the maximum value of the performance coefficient.

9. The control method according to claim 2, wherein, The step of determining the control sequence for controlling the chiller based on load forecast results and partial load factor constraints includes: Based on the cooling load demand prediction curve, multiple initial control sequences are generated, each initial control sequence being limited by a minimum on / off interval; and The control sequence is determined from multiple initial control sequences based on partial load rate constraints.

10. A control system for controlling a chiller in a chiller unit, comprising: The load forecasting module is configured to predict the load of the chiller based on time characteristics, weather characteristics, and historical cooling load characteristics of the chiller. The filtering module is configured to determine a control sequence for controlling the chillers based on load forecast results and partial load factor constraints. The control sequence includes the number of chillers to be turned on and their operating order in the chiller unit. The optimization module is configured to select the optimal control sequence from the control sequence based on the chiller model; as well as The control module is configured to control the operation of the chiller based on the optimal control sequence.

11. The control system according to claim 10, wherein, The load forecasting module is specifically configured as follows: The XGBoost model receives the time characteristics, weather characteristics, and historical cooling load characteristics of the chiller, outputs a cooling load demand prediction curve, and predicts the load of the chiller based on the prediction curve.

12. The control system according to claim 11, wherein, The optimization module also includes a chiller model trainer, configured to determine target feature parameters based on initial feature parameters related to the chiller, and in response to the artificial neural network receiving the target feature parameters, to build the chiller model and output the chiller's energy consumption and coefficient of performance. The initial characteristic parameters include at least one of the following: chiller operation list, chilled water flow rate, cooling water flow rate, chiller unit temperature point data, outdoor ambient temperature, total cooling load, and partial load rate.

13. The control system according to claim 12, wherein, The chiller model trainer is specifically configured to use statistical methods of covariance matrix and confidence ellipse to determine the effective range of the initial feature parameters, and to determine the initial feature parameters that are within the effective range as the target feature parameters.

14. The control system according to claim 10, wherein, The optimization module is specifically configured to determine the energy consumption of the chiller unit corresponding to the control sequence within a specific time interval based on the chiller model, calculate the energy consumption of each control process in any control sequence based on the chiller model, and select the optimal chiller control sequence based on the comparison results of the energy consumption of multiple chiller units and the energy consumption of each control step in the control sequence that controls the operation of the chiller unit.

15. The control system according to claim 14, wherein, The optimization module is further configured to establish a state transition equation to determine a control sequence and the cumulative energy consumption of the control sequence from the start time to time t, wherein a control action in the control sequence determines the on or off state of the chiller at the next time moment, where t is a positive integer, and to store the tuple formed by the control sequence and the cumulative energy consumption if the tuple is not stored.

16. The control system according to claim 14, wherein, The optimization module is further configured to search for a control sequence that satisfies the cooling load and operating constraints when the partial load rate meets the first condition, and to search for a control sequence that satisfies the cooling load and operating constraints when the partial load rate meets the second condition if the control sequence is not found; wherein the numerical range of the second condition is greater than the numerical range of the first condition, and the numerical range of the first condition is the range in which the partial load rate is close to the maximum value of the performance coefficient.

17. The control system according to claim 10, wherein, The filtering module is specifically configured to generate multiple initial control sequences based on the cooling load demand prediction curve. The initial control sequences are subject to minimum on and off intervals and are determined from the multiple initial control sequences based on partial load rate constraints.

18. An electronic device comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the chiller control method according to any one of claims 1-9.

19. A non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform the chiller control method according to any one of claims 1-9.

20. A computer program product comprising a computer program that, when executed by a processor, implements the steps of the chiller control method according to any one of claims 1-9.

Images