Refrigeration host parallel system prediction and energy efficiency control method and system

Abstract

The invention discloses a refrigeration host parallel system prediction and energy efficiency control method and system, relates to the technical field of refrigeration system energy efficiency optimization control, and is used for solving the problem that energy efficiency optimization and load dynamic response are difficult to cooperate when a plurality of refrigeration hosts are in parallel operation under variable working conditions. According to the method, a working condition-load-energy efficiency mapping relation is established by constructing a refrigeration host three-dimensional energy efficiency mapping table, day-ahead and intra-day load prediction is executed in parallel based on historical data and environmental parameters, and a comprehensive load value is obtained through prediction error inverse proportion dynamic weighted fusion. And enumerating an optimal host operation combination and issuing a control instruction by taking the maximization of the comprehensive revenue function as a target, and updating prediction model parameters on line by utilizing a residual error of an actual load and a prediction value. The method is mainly used for energy efficiency optimization and self-adaptive control of a central air conditioning system, a ground source heat pump system or an industrial refrigerating unit group, the overall energy efficiency of the system can be effectively improved, and frequent start and stop are reduced.

Description

Technical Field

[0001] This invention relates to the field of energy efficiency optimization and control technology for refrigeration systems. More specifically, this invention relates to a predictive and energy efficiency control method and system for a parallel refrigeration unit system. Background Technology

[0002] With the rapid development of large public buildings and industrial facilities, their energy consumption problems have become increasingly prominent, among which the refrigeration units of central air conditioning systems are the main energy-consuming units. To meet the fluctuating cooling load demand, the parallel operation of multiple refrigeration units has become the standard configuration, and their energy efficiency level directly affects the operating economy and carbon emissions of the entire system. Therefore, how to accurately optimize and control the energy efficiency of parallel refrigeration unit systems has always been one of the core research topics in the field of building energy conservation.

[0003] Traditional control strategies often rely on simple start-stop logic or empirical formulas based on fixed operating points, such as starting and stopping the chiller based on return water temperature or pressure parameters. These methods have significant limitations because they treat the chiller's coefficient of performance (COP) as a constant value or solely dependent on the load rate, severely neglecting the substantial impact of variable operating parameters such as cooling water temperature and ambient humidity on the actual energy efficiency of the chiller. In actual operation, the same chiller can exhibit drastically different energy efficiency performance under the same load rate at different cooling water inlet temperatures, making it difficult for traditional methods to achieve true energy efficiency optimization in real-world, variable environments.

[0004] Furthermore, system load itself exhibits strong time-varying and uncertainties, influenced by multiple factors such as weather, human activity, and scheduling. Although existing research has attempted to introduce load forecasting techniques, a single forecasting model often struggles to simultaneously guarantee the accuracy of capturing long-term trends and responding to short-term fluctuations. There may be discrepancies between day-ahead and intraday forecasts. Without an effective fusion and reliability assessment mechanism, directly using erroneous forecasts for optimization control could lead to decision-making errors, or even cause frequent equipment start-ups and shutdowns, damaging equipment lifespan.

[0005] Furthermore, existing optimization control objectives are often too singular, either pursuing the highest instantaneous energy efficiency or the lowest load tracking deviation, failing to incorporate multiple conflicting objectives such as energy efficiency gains, load matching accuracy, and equipment switching costs into a unified framework for comprehensive consideration. A slightly more energy-efficient operating configuration, if requiring the start-up and shutdown of multiple main units, may have its energy consumption and mechanical losses during switching processes offset or even exceed its energy efficiency gains. Such multi-dimensional trade-offs are difficult to address using traditional methods.

[0006] In summary, existing technologies for controlling parallel refrigeration systems generally suffer from three major shortcomings: first, they fail to accurately characterize the dynamic impact of varying operating conditions on the system's performance; second, they fail to effectively integrate and utilize predictive information across multiple time scales and assess its reliability; and third, they lack a comprehensive optimization decision-making mechanism that can coordinate energy efficiency, load tracking, and equipment stability. These shortcomings lead to the system operating under suboptimal conditions for extended periods, resulting in significant energy waste. Therefore, developing a predictive and energy-efficient control method that overcomes these challenges has extremely important practical significance and application value. Summary of the Invention

[0007] One object of the present invention is to solve at least the above-mentioned problems and to provide at least the advantages that will be described later.

[0008] Another objective of this invention is to provide a predictive and energy efficiency control method for a parallel cooling system. This method accurately characterizes the dynamic energy efficiency characteristics of the cooling system, integrates high-reliability load prediction across multiple time scales, and comprehensively optimizes energy efficiency, load tracking, and equipment stability. This significantly improves the overall energy efficiency of the system, effectively avoids frequent start-ups and shutdowns of the cooling system, and reduces operating costs and equipment wear.

[0009] To achieve these objectives and other advantages according to the present invention, a predictive and energy efficiency control method for a parallel refrigeration unit system is provided, comprising: S1. Construct a three-dimensional energy efficiency mapping table for each refrigeration unit. The three-dimensional energy efficiency mapping table takes the operating condition parameters and load rate of the refrigeration unit as input variables and the real-time energy efficiency value COP as output variable. S2. Based on historical load data and environmental parameters, execute the day-ahead periodic forecasting model and the intraday rolling forecasting model in parallel to generate day-ahead load forecasts and intraday load forecasts, respectively. S3. Based on the error between the day-ahead load forecast value and the intraday load forecast value and the actual load value in step S2, dynamically calculate the reliability coefficient of the two load forecast values, and perform weighted fusion of the two load forecast values ​​based on the reliability coefficient to generate a comprehensive load forecast value. S4. Using the comprehensive load forecast value obtained in step S3 as the expected output, enumerate all feasible chiller operation combinations. For each chiller operation combination, query the three-dimensional energy efficiency mapping table in step S1 to obtain the real-time energy efficiency value of each chiller, construct a comprehensive benefit function, determine the optimal chiller operation combination with the optimization objective of maximizing the comprehensive benefit function, and generate and issue control commands for loading or unloading the chillers. S5. Collect the actual load value, calculate the residuals between it and the day-ahead load forecast value and the intraday load forecast value, and use the residuals to update the parameters of the day-ahead cycle forecast model and the intraday rolling forecast model in step S2 online, and automatically correct the day-ahead cycle forecast model and the intraday rolling forecast model for subsequent forecast cycles. The comprehensive revenue function is at least related to the system's average energy efficiency level, the deviation level between the total cooling capacity and the comprehensive load forecast, and the number of times the refrigeration unit needs to be started and stopped.

[0010] Preferably, the comprehensive benefit function in step S4 is: The optimization objective is: ; Among them, F i Let be the overall return value of the i-th operating combination, which is dimensionless; Let be the normalized average energy efficiency value for the i-th operating combination, which is dimensionless; COP i The average energy efficiency (COP) for the i-th operating combination; ref Take the standard COP value of the refrigeration unit under rated operating conditions; Q i The total cooling capacity of the i-th operating combination is expressed in kW. L pred The total load forecast is in kW. This is the normalized value of the deviation between the total cooling capacity and the predicted comprehensive load for the i-th operating combination, and is dimensionless. Q ref The normalized baseline value is the rated total cooling capacity of the system, in kW. The normalized value of the number of refrigeration unit start-ups and shutdowns required to switch from the current state to the i-th combination; N represents the number of times the cooling unit is started and stopped when switching from the current state to the i-th combination; N ref Take the preset maximum number of start-stop cycles per session; ω1, ω2, and ω3 are the weight coefficients of each item, which are dimensionless and obtained through an optimization algorithm based on historical operating data. ω1 + ω2 + ω3 = 1.

[0011] Preferably, the day-ahead cycle load model in step S2 adopts the Holt-Winters additive model to predict the load profile for the next 24-hour cycle, specifically including: Initialize the level component, trend component, and seasonal component of the Holt-Winters additive model based on historical load time series data; For each moment t, dynamically update the level component, trend component, and seasonal component of the Holt-Winters additive model according to the latest monitored load value using the following recurrence formula: Level component: ; Trend component: ; Seasonal component: ; Generate the load prediction values for the next k steps based on the updated level component, trend component, and seasonal component of the Holt-Winters additive model as the day-ahead load prediction values: Load prediction values for the next k steps ; where l t and l t-1 are the level components of the model, which are the estimated load values at the current moment t and the previous moment t - 1, in kW; y t is the actual load measurement value at the current moment t, in kW; b t and b t-1 are the trend components of the model, which are the load change amounts at the current moment t and the previous moment t - 1, in kW; s t and s t-p are the seasonal components of the model, which are the seasonal load deviation values at the current moment t and the moment t one cycle p ago, in kW; p is the cycle length, 24 h; γ, ε, and μ are the level smoothing coefficient, trend smoothing coefficient, and seasonal smoothing coefficient respectively. γ ranges from 0.1 to 0.3, dimensionless; ε ranges from 0.01 to 0.2, dimensionless; μ ranges from 0.2 to 0.5, dimensionless; k is the prediction step size, k < p, dimensionless; is the load prediction value at the moment t + k, in kW; s t-p+k is the seasonal load deviation value when going back one cycle p from the moment t + k, in kW.

[0012] Preferably, the intra-day rolling prediction model in step S2 is constructed based on multiple linear regression and is used to achieve future short-term fine load prediction, specifically including: Construct an intra-day rolling prediction model based on the latest monitored historical load data, historical prediction residuals, and real-time meteorological parameters: ; Among them, L i,t The current intraday load forecast value at time t, in kW, is output by the intraday rolling forecast model. θ0 is a constant in the intraday rolling forecast model, in kW; L t-j The measured historical load value in kW is the value of the j-th hour prior to the current time t. θ j The regression coefficients for the historical load term are dimensionless. m represents the number of historical load items considered in the intraday rolling forecast model, which is dimensionless. B t-q kW represents the difference between the historical predicted value and the measured value q hours before the current time t. A q The regression coefficients of the historical prediction residuals are dimensionless. n is the number of historical forecast residual items considered in the intraday rolling forecast model, which is dimensionless; D1 is the regression coefficient of the outdoor temperature term, kW / ℃; T out Let t be the outdoor dry-bulb temperature monitored at the current time, in °C; D2 is the regression coefficient for the outdoor relative humidity term, kW / %; RH represents the outdoor relative humidity monitored at time t, in %.

[0013] Preferably, the specific implementation process of weighted fusion of the two load forecast values ​​to generate a comprehensive load forecast value in step S3 includes: Based on the latest forecast errors of the day-ahead cycle forecast model and the intraday rolling forecast model, the reliability coefficients of the day-ahead load forecast and the intraday load forecast are calculated using the following method: ; R2 = 1 - R1; The two load forecasts, namely the day-ahead load forecast and the intraday load forecast, are weighted and fused to generate a composite load forecast value L. pred : ; Wherein, R1 is the confidence coefficient of the day-ahead load forecast, with a value of 0 to 1; R2 is the confidence coefficient of the intraday load forecast, with a value of 0 to 1; L1 is the daytime load forecast value obtained in step S2, in kW; L2 is the daily load forecast value obtained in step S2, in kW; L actual The latest actual load measurement, in kW; c is the smoothing constant, which is set to 10. -6 kW; L pred This is the comprehensive load forecast value, in kW.

[0014] Preferably, the method for obtaining the weight coefficients ω1, ω2, and ω3 of the comprehensive benefit function in step S4 includes: Based on historical meteorological data and load datasets, corresponding refrigeration unit operation records are used to construct a training sample set containing various operating conditions. Under the constraint that ω1+ω2+ω3=1 and all terms are greater than or equal to 0, generate multiple sets of candidate weight coefficient combinations; Using the training sample set as input, the prediction and energy efficiency control method of the parallel system of the refrigeration unit is simulated to obtain the simulated operation command sequence and the corresponding total cooling capacity, actual load and equipment start-up and shutdown records of the system. Based on simulation records, construct long-term comprehensive performance indices corresponding to weighting coefficients ω1, ω2, and ω3: ; The combination of weight coefficients ω1, ω2, and ω3 that minimizes the long-term comprehensive performance index can be searched using a genetic algorithm or a particle swarm optimization algorithm. Among them, E total The total energy consumption of all cooling units in the system during the training period, in kW·h; D total The integral of the absolute value of the deviation between the total cooling capacity of the system and the actual load at each moment during the training period, expressed in kW·h. S total This is the sum of the start and stop times of all refrigeration units during the training period; λ1 is D total Conversion factor, value 1~10, dimensionless; λ2 is S total Conversion factor, ranging from 1 to 50; kW·h / time.

[0015] The present invention further claims protection for the control system of the predictive and energy efficiency control method for the parallel system of the refrigeration unit, comprising: The energy efficiency mapping table construction module is used to construct a three-dimensional energy efficiency mapping table to characterize the performance of each refrigeration unit under varying operating conditions. The three-dimensional energy efficiency mapping table takes the real-time inlet temperature and outlet temperature of the cooling medium of the refrigeration unit and the load rate of the refrigeration unit as input variables, and its corresponding real-time energy efficiency value COP as output variables. The load forecasting module is used to execute the day-ahead periodic forecasting model and the intraday rolling forecasting model in parallel based on historical load data and multi-source environmental parameters, and generate day-ahead load forecasting values ​​and intraday load forecasting values ​​respectively. The multi-source environmental parameters include at least outdoor dry-bulb temperature, wet-bulb temperature and relative humidity. The credibility fusion module is used to introduce a credibility dynamic allocation mechanism based on the inverse ratio of prediction error, calculate the credibility coefficient of the day-ahead load forecast and the intraday load forecast in real time, and perform weighted fusion of the two load forecasts according to the credibility to generate a comprehensive load forecast. The credibility coefficient is dynamically adjusted according to the absolute error between the load forecast and the actual load in the previous period. The optimization control module is used to enumerate all feasible chiller operation combinations with the comprehensive load forecast value as the expected output. For each chiller operation combination, it queries the three-dimensional energy efficiency mapping table to obtain the real-time energy efficiency value of each chiller, calculates the overall energy efficiency of the combined system, constructs a comprehensive benefit function, and solves the maximum value of the comprehensive benefit function to dynamically determine the optimal number of chillers to operate and the combination strategy. It compares the current chiller operation combination with the optimal chiller operation combination, generates and issues instructions to load or unload the chillers. The model update module is used to collect actual load values, calculate the residuals between them and the day-ahead load forecast values ​​and the intraday load forecast values, and use the residuals to update the parameters of the day-ahead periodic forecast model and the intraday rolling forecast model online through a recursive algorithm, and automatically correct the day-ahead periodic forecast model and the intraday rolling forecast model for subsequent forecast periods.

[0016] The present invention has at least the following beneficial effects: Firstly, by constructing a detailed three-dimensional energy efficiency table, this invention establishes a dynamic correspondence between cooling water temperature, load rate, and chiller energy efficiency, breaking through the limitation of traditional methods that treat energy efficiency as a fixed value. It accurately reflects the true performance of the chiller under actual operating conditions, providing a precise energy efficiency data basis for system optimization and significantly improving the accuracy and reliability of control decisions. Secondly, this invention employs a parallel day-ahead periodic forecasting and intraday rolling forecasting model, combined with multi-source environmental parameters for load forecasting. This not only captures long-term load trends but also achieves precise tracking of short-term fluctuations. Furthermore, through a dynamic reliability weighted fusion mechanism, it effectively improves the accuracy and robustness of the comprehensive load forecast value, providing high-quality input for subsequent optimization control. Thirdly, this invention constructs a comprehensive benefit function with system energy efficiency, load tracking deviation and equipment switching frequency as the core, and uses an enumeration optimization strategy to dynamically determine the optimal host operation combination. This can effectively balance energy efficiency improvement and equipment loss while ensuring efficient system operation, and achieve a unified optimization of economy and stability. Fourth, this invention utilizes the prediction residuals to update the prediction model parameters online in real time through a recursive algorithm with a forgetting factor, enabling the prediction model to continuously learn and adaptively adjust. This allows it to continuously track changes in system characteristics and environmental fluctuations, significantly improving the model's long-term prediction accuracy and the system's overall adaptability. This effect corresponds to the online model parameter update mechanism in claim 5. Fifth, the predictive and energy efficiency control method for parallel refrigeration systems provided by this invention is integrated into a control system, forming a closed-loop automatic control from data perception, prediction, decision-making to execution. This not only significantly improves the automation and energy efficiency of the refrigeration system operation, but also reduces the reliance on human experience, providing a highly efficient, reliable and adaptive comprehensive solution for energy-saving operation of large buildings.

[0017] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the prediction and energy efficiency control method for a parallel cooling system according to one technical solution of the present invention; Figure 2 This is a flowchart illustrating the day-ahead cycle prediction model described in another technical solution of the present invention; Figure 3 This is a flowchart illustrating the intraday rolling forecast model described in another technical solution of the present invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0020] It should be understood that terms such as “having,” “comprising,” and “including” as used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0021] like Figures 1-3 As shown, the present invention provides a method for predictive and energy efficiency control of a parallel refrigeration unit system, comprising: S1. Construct a three-dimensional energy efficiency mapping table for each refrigeration unit. The three-dimensional energy efficiency mapping table takes the operating condition parameters and load rate of the refrigeration unit as input variables and the real-time energy efficiency value COP as output variable. S2. Based on historical load data and environmental parameters, execute the day-ahead periodic forecasting model and the intraday rolling forecasting model in parallel to generate day-ahead load forecasts and intraday load forecasts, respectively. S3. Based on the error between the day-ahead load forecast value and the intraday load forecast value and the actual load value in step S2, dynamically calculate the reliability coefficient of the two load forecast values, and perform weighted fusion of the two load forecast values ​​based on the reliability coefficient to generate a comprehensive load forecast value. S4. Using the comprehensive load forecast value obtained in step S3 as the expected output, enumerate all feasible chiller operation combinations. For each chiller operation combination, query the three-dimensional energy efficiency mapping table in step S1 to obtain the real-time energy efficiency value of each chiller, construct a comprehensive benefit function, determine the optimal chiller operation combination with the optimization objective of maximizing the comprehensive benefit function, and generate and issue control commands for loading or unloading the chillers. S5. Collect the actual load value, calculate the residuals between it and the day-ahead load forecast value and the intraday load forecast value, and use the residuals to update the parameters of the day-ahead cycle forecast model and the intraday rolling forecast model in step S2 online, and automatically correct the day-ahead cycle forecast model and the intraday rolling forecast model for subsequent forecast cycles. The comprehensive revenue function is at least related to the system's average energy efficiency level, the deviation level between the total cooling capacity and the comprehensive load forecast, and the number of times the refrigeration unit needs to be started and stopped.

[0022] The above technical solution achieves energy efficiency control of a parallel chiller system by constructing a three-dimensional energy efficiency mapping table, integrating load forecasts across multiple time scales, dynamically optimizing the chiller unit operation combination, and using an online correction prediction model. The system establishes a three-dimensional energy efficiency mapping table for each chiller unit, using the inlet temperature (5~40℃), outlet temperature (7~45℃), and load rate (10%~100%) of the cooling medium as inputs, and real-time COP (1.0~6.0) as the output. A parallel prediction architecture is adopted: the day-ahead cycle prediction model uses historical load data and environmental parameters (outdoor dry-bulb temperature, wet-bulb temperature, relative humidity) to predict the load profile for the next 24 hours using the Holt-Winters model; the intraday rolling prediction model uses short-term data (e.g., sampling every 15~60 minutes) and a multi-source linear regression model for precise predictions over a short period. Furthermore, a dynamic reliability allocation mechanism integrates the results of the two types of predictions (day-ahead cycle prediction and intraday rolling prediction), where the error calculation period is typically 1 hour, but can also be set according to the building's energy consumption. Then, using the integrated load forecast as the target, all chiller combinations are enumerated (e.g., 4 chillers can produce 15 combinations). The COP of each chiller is obtained by querying the three-dimensional energy efficiency mapping table, the overall system energy efficiency is calculated (using a weighted average), and a comprehensive revenue function is constructed (the weighting coefficients ω1, ω2, and ω3 are determined through optimization using historical data). Finally, by comparing the current state with the optimal combination, start-stop commands are generated, and the prediction model parameters are updated online using the residuals between the actual load and the predicted value (either recursive least squares or gradient descent algorithm can be used).

[0023] Based on the above technical solution, taking a parallel system of building refrigeration units as an example, its specific workflow begins with the data acquisition phase. Temperature sensors (PT1000 platinum resistance thermometers can be selected) and electricity meters installed at the inlet and outlet of the cooling water pipes collect real-time operating data of the refrigeration units. Outdoor parameters are collected through environmental sensors (Visa temperature and humidity integrated probes can be selected). The system, deployed on a cloud platform or local server, performs daily periodic forecasting at 00:00 and intraday rolling forecasting every hour, generating a comprehensive load value through reliability calculation. The system's control module performs optimization calculations every 30 minutes, enumerating all combinations of refrigeration units based on the comprehensive load forecast value, calculating the combination of refrigeration units with the highest benefit, and issuing start / stop commands according to this combination. After execution, the system continues to collect actual load data and calculate the residual between the actual load and the forecast value, updating the parameters of the daily periodic forecasting model and the intraday rolling forecasting model with the residual, thus achieving closed-loop control.

[0024] According to the above technical solution, it firstly significantly improves system energy efficiency by fusing a three-dimensional energy efficiency mapping table with multi-timescale prediction. Experiments show that the energy saving rate can reach 15-25% (based on the test results of a certain building), avoiding the energy efficiency estimation deviation caused by the variable operating conditions of traditional methods. Secondly, it enhances the system's adaptability through a dynamic reliability mechanism and online model correction, especially improving prediction accuracy during sudden weather changes or drastic load fluctuations, such as sudden temperature drops during summer thunderstorms. Finally, the comprehensive benefit function designed in this invention coordinates energy efficiency, load matching, and equipment switching losses, preventing frequent start-ups and shutdowns of the chiller and extending its lifespan, while ensuring accurate fulfillment of load demands. It is particularly suitable for scenarios with strict temperature control requirements, such as commercial complexes and data centers. In one technical solution, the comprehensive benefit function in step S4 is: The optimization objective is: ; Among them, F i Let be the overall return value of the i-th operating combination, which is dimensionless; Let be the normalized average energy efficiency value for the i-th operating combination, which is dimensionless; COP i The average energy efficiency (COP) for the i-th operating combination; ref Take the standard COP value of the refrigeration unit under rated operating conditions; Q i The total cooling capacity of the i-th operating combination is expressed in kW. L pred The total load forecast is in kW. This is the normalized value of the deviation between the total cooling capacity and the predicted comprehensive load for the i-th operating combination, and is dimensionless. Q ref The normalized baseline value is the rated total cooling capacity of the system, in kW. The normalized value of the number of refrigeration unit start-ups and shutdowns required to switch from the current state to the i-th combination; N represents the number of times the cooling unit is started and stopped when switching from the current state to the i-th combination; N ref Take the preset maximum number of start-stop cycles per session; ω1, ω2, and ω3 are the weight coefficients of each item, which are dimensionless and obtained through an optimization algorithm based on historical operating data. ω1 + ω2 + ω3 = 1.

[0025] This invention further optimizes the specific formula of the comprehensive benefit function. Based on a parallel cooling system, it establishes for the first time three core objectives that have a significant impact and are mutually competitive: highest energy efficiency, most accurate load tracking, and most stable system. Highest energy efficiency aims for the most energy-efficient system operation; most accurate load tracking aims for the best match between the total cooling capacity of the cooling units and the actual needs of the building; and most stable system aims for the fewest equipment start-stop switching times. In this application, energy efficiency level is represented by the normalized average energy efficiency value, load tracking accuracy is represented by absolute deviation, and system stability is represented by the number of start-stop cycles. When constructing the multi-objective unified framework, this invention identifies inherent conflicts among the three objectives. To resolve these conflicts, a weighted summation method is used to unify them into a comprehensive framework. Due to the different dimensions and orders of magnitude of the indicators, they are first normalized to become dimensionless scalars, and then weighted to obtain the final summation benefit function. This invention is the first to incorporate three conflicting core objectives into a unified, quantifiable decision-making framework. It simultaneously weighs energy efficiency, tracking accuracy, and equipment wear through a mathematical expression. The negative impact of equipment start-up and shutdown is usually only considered as a constraint. This invention is the first to quantify it and incorporate it directly into the comprehensive benefit function as a core optimization objective. This significantly improves the practicality and economy of the system, extends the service life of the chiller, and effectively solves the fundamental technical problem of the difficulty in coordinating energy efficiency optimization and stable operation of parallel chiller systems.

[0026] like Figure 2 In one technical solution, the day-ahead cycle load model in step S2 adopts the Holt-Winters additive model to predict the load profile for a future 24-hour cycle, specifically including: Based on historical load time series data, the horizontal, trend, and seasonal components of the Holt-Winters additive model were initialized. For each moment \(t\), the level component, trend component, and seasonal component of the Holt-Winters additive model are dynamically updated according to the latest monitored load value using the following recursive formula: Level component: ; Trend component: ; Seasonal component: ; Based on the updated level component, trend component, and seasonal component of the Holt-Winters additive model, the load prediction value for the next \(k\) steps is generated as the day-ahead load prediction value: Load prediction value for the next \(k\) steps ; where \(l\) t and \(l\) t-1 are the level components of the model, which are the estimated load values at the current moment \(t\) and the previous moment \(t - 1\), in kW; \(y\) t is the actual load measurement value at the current moment \(t\), in kW; \(b\) t and \(b\) t-1 are the trend components of the model, which are the load change amounts at the current moment \(t\) and the previous moment \(t - 1\), in kW; \(s\) t and \(s\) t-p are the seasonal components of the model, which are the seasonal load deviation values at the current moment \(t\) and the moment \(t\) one cycle \(p\) ago, in kW; \(p\) is the cycle length, 24 h; \(\gamma\), \(\varepsilon\), and \(\mu\) are the level smoothing coefficient, trend smoothing coefficient, and seasonal smoothing coefficient respectively. \(\gamma\) ranges from 0.1 to 0.3, dimensionless; \(\varepsilon\) ranges from 0.01 to 0.2, dimensionless; \(\mu\) ranges from 0.2 to 0.5, dimensionless; \(k\) is the prediction step, \(k < p\), dimensionless; is the load prediction value at the moment \(t + k\), in kW; \(s\) t-p+k is the seasonal load deviation value when going back one cycle \(p\) from the moment \(t + k\), in kW.

[0027] The aforementioned technical solution further optimizes the day-ahead cyclical load forecasting model by dynamically updating the horizontal, trend, and seasonal components using the Holt-Winters additive model to capture the cyclical variation patterns of the load. Specifically, the system first initializes the horizontal component (representing the baseline load value), trend component (representing the rate of load change), and seasonal component (representing the cyclical deviation of the same time period each day) of the Holt-Winters additive model based on historical 24-hour load time series data. The Holt-Winters additive model uses three smoothing coefficients for dynamic adjustment. The horizontal smoothing coefficient γ typically ranges from 0.1 to 0.3 to balance the latest data with historical horizontal values; the trend smoothing coefficient ε ranges from 0.01 to 0.2 to control the sensitivity of trend changes; and the seasonal smoothing coefficient μ ranges from 0.2 to 0.5 to adapt to daily cyclical fluctuations. These smoothing coefficients are determined through training and optimization using historical data and embedded into the forecasting algorithm deployed on an industrial computer (local server) or cloud platform. During forecasting, each time the model receives a new actual load measurement, it updates three components using the published recursive formula: the horizontal component combines the latest load value with seasonal adjustments, the trend component corrects the direction of change, and the seasonal component updates periodic deviations. The updated model then overlays the horizontal value, the trend extension value, and the corresponding seasonal component to generate hourly load forecasts for the next 24 hours (forecast step size k=1~24). The environmental parameters relied upon by the Holt-Winters additive model are provided by outdoor sensors.

[0028] This invention further uses a parallel cooling system of a commercial building as an example to illustrate the specific workflow of using the Holt-Winters additive model to achieve day-ahead periodic load forecasting. The forecasting system retrieves historical load data from the system database for the past week (adjustable according to settings) and initializes the Holt-Winters model parameters. Every hour on the hour (for example, 10:00 AM), the system collects the latest actual load value and environmental parameters, calculates the current seasonal component deviation (compared with the data from yesterday's 10:00 AM), and then updates the horizontal component, trend component, and seasonal component according to the recursive formula. After the update, the model generates a 24-hour forecast (forecast value up to 10:00 AM tomorrow) and outputs it to the fusion module for fusion with the intraday rolling forecast model. If the weather changes abruptly (such as afternoon downpours), the system continuously corrects the components based on the actual values ​​to ensure forecast adaptability.

[0029] According to the above technical solution, the Holt-Winters additive model and recursive formula provided by this invention significantly improve the ability of day-ahead cycle forecasting to capture daily cycle patterns by explicitly modeling trends and seasonal components. It is particularly suitable for scenarios with fixed operating hours, such as office buildings and shopping malls, and the prediction error is significantly reduced compared to traditional linear models. The dynamic smoothing coefficient mechanism enables the model to quickly respond to load changes, such as sudden temperature increases or decreases in summer, by adjusting coefficient weights to reduce prediction lag. The entire model has high computational efficiency and can run in real time on ordinary industrial controllers without requiring expensive hardware, providing a reliable timescale basis for subsequent energy efficiency optimization.

[0030] like Figure 3 In one technical solution, the intraday rolling forecast model described in step S2 is constructed based on multiple linear regression to achieve precise short-term load forecasting, specifically including: Based on the latest monitored historical load data, historical forecast residuals, and real-time meteorological parameters, an intraday rolling forecast model is constructed: ; Among them, L i,t The current intraday load forecast value at time t, in kW, is output by the intraday rolling forecast model. θ0 is a constant in the intraday rolling forecast model, in kW; L t-j The measured historical load value in kW is the value of the j-th hour prior to the current time t. θ j The regression coefficients for the historical load term are dimensionless. m represents the number of historical load items considered in the intraday rolling forecast model, which is dimensionless. B t-q kW represents the difference between the historical predicted value and the measured value q hours before the current time t. A q The regression coefficients of the historical prediction residuals are dimensionless. n is the number of historical forecast residual items considered in the intraday rolling forecast model, which is dimensionless; D1 is the regression coefficient of the outdoor temperature term, kW / ℃; T out Let t be the outdoor dry-bulb temperature monitored at the current time, in °C; D2 is the regression coefficient for the outdoor relative humidity term, kW / %; RH represents the outdoor relative humidity monitored at time t, in %.

[0031] The above technical solution achieves high-precision short-term load forecasting by constructing a multiple linear regression model. The intraday rolling forecasting model provided by this invention comprehensively considers the multiple influences of historical load values, historical forecast deviations, and real-time air parameters. Specifically, the intraday rolling forecasting model provided by this invention uses measured load data from the past few hours (for better understanding, taking the load of the previous 4 hours as an example, i.e., m=4), the residual between recent forecast values ​​and actual values ​​(taking the residual of the previous 2 hours as an example, i.e., n=2), and the current outdoor dry-bulb temperature and relative humidity as input features. Regression coefficient θ j A q D1 and D2 are determined through training on historical data. During training, standardized Z-score normalization can be used to process the feature data, and L2 regularization is introduced to prevent overfitting. The intraday rolling forecast model is deployed on a local computer or cloud platform, and performs a forecast every 15 minutes. Real-time acquisition of meteorological sensor data ensures the timeliness of the input.

[0032] This invention further uses a parallel cooling system of a commercial building as an example to specifically illustrate the implementation process of intraday rolling forecasting using a multiple linear regression model. The workflow of the intraday rolling forecasting model begins 10 minutes before the hour (which can be adjusted according to the actual scenario). The system reads the measured load values ​​of the most recent 4 hours, the prediction residuals of the previous 2 hours, and the current meteorological data from the database. The intraday rolling forecasting model calls the pre-trained regression coefficients to calculate and outputs the load forecast values ​​for the next 15 minutes. If there is a deviation between the actual load and the predicted value, the deviation is recorded and used for the online update of the model coefficients in the next cycle (such as adjusting the regression coefficients by recursive least squares). At the same time, the forecast results are transmitted to the credibility fusion module and weighted with the previous day's load forecast values.

[0033] According to the above technical solution, the intraday rolling forecast model provided by this invention significantly improves the short-term forecast's responsiveness to sudden weather changes by integrating multi-source dynamic parameters, especially real-time meteorological data. For example, in the case of a sudden drop in temperature caused by a thunderstorm in the afternoon during summer, the forecast error can be controlled within 5%. The introduction of historical residual terms effectively corrects system biases and avoids error accumulation. For example, if the load is continuously overestimated in the morning, the afternoon forecast automatically lowers the baseline value, thereby improving the stability of the intraday rolling forecast model. The intraday rolling forecast model provided by this invention is based on a multiple linear regression model, which has a simple structure, low computational load, and can run on low-cost embedded devices. It provides high-frequency and accurate load commands to the refrigeration unit, ensuring the real-time performance of the energy efficiency optimization strategy.

[0034] In one of the technical solutions, the specific implementation process of weighted fusion of the two load forecast values ​​to generate a comprehensive load forecast value in step S3 includes: Based on the latest forecast errors of the day-ahead cycle forecast model and the intraday rolling forecast model, the reliability coefficients of the day-ahead load forecast and the intraday load forecast are calculated using the following method: ; R2 = 1 - R1; The two load forecasts, namely the day-ahead load forecast and the intraday load forecast, are weighted and fused to generate a composite load forecast value L. pred : ; Wherein, R1 is the confidence coefficient of the day-ahead load forecast, with a value of 0 to 1; R2 is the confidence coefficient of the intraday load forecast, with a value of 0 to 1; L1 is the daytime load forecast value obtained in step S2, in kW; L2 is the daily load forecast value obtained in step S2, in kW; L actual The latest actual load measurement, in kW; c is the smoothing constant, which is set to 10. -6 kW; L pred This is the comprehensive load forecast value, in kW.

[0035] The aforementioned technical solution further achieves intelligent fusion of day-ahead and intraday load forecasts by optimizing the dynamic reliability allocation mechanism. This solution automatically adjusts the weights of the day-ahead periodic forecast model and the intraday rolling forecast model based on their latest performance. Specifically, the system acquires the latest day-ahead forecast value L1 and intraday forecast value L2, compares them with the latest collected real-time load value, calculates their respective absolute errors, and then uses a weighted algorithm based on inverse error proportion. The forecast model with the smaller error obtains a higher reliability coefficient, with a value ranging from 0 to 1. To prevent calculation anomalies when the error is extremely small, this invention introduces a very small smoothing constant c, with a value of 10, during the calculation. -6 In actual operation, the value of c can be adjusted, and the smoothing constant c ensures the stability of the value. The resulting comprehensive load forecast is obtained by weighting and summing the two forecasts according to their respective confidence coefficients, thus taking into account both the advantages of long-term trend prediction and short-term fluctuation response.

[0036] According to the above technical solution, the dynamic weighting mechanism enables the system to adaptively trust more accurate prediction sources, significantly improving the accuracy of the final load forecast. Especially when sudden weather changes cause a single model to fail, the fusion prediction error can be significantly reduced. The introduction of a smoothing constant avoids the instability of mathematical calculations, ensuring that the system can still reasonably allocate weights and maintain the continuity of the output results even when the prediction error is close to zero. The fusion method provided by this invention enhances the robustness of the overall prediction, utilizing both the day-ahead cycle prediction model's grasp of cyclical patterns and the intraday rolling prediction model's sensitivity to real-time changes, providing a more reliable decision-making basis for subsequent energy efficiency optimization.

[0037] In one technical solution, step S5 utilizes the residuals to update the parameters of the day-ahead cycle forecasting model and the intraday rolling forecasting model from step S2 online using a recursive algorithm. Specifically, an online learning and parameter update of the day-ahead cycle forecasting model and the intraday rolling forecasting model is achieved through a recursive algorithm with a forgetting factor. Real-time acquisition of actual load values ​​is used, and comparison with the comprehensive forecast values ​​generates forecast residuals. These residuals are then used to dynamically adjust the internal parameters of the day-ahead cycle forecasting model and the intraday rolling forecasting model, respectively. For the day-ahead cycle forecasting model (Holt-Winters), the system recursively corrects its horizontal, trend, and seasonal components. The correction magnitude is controlled by a forgetting factor (typically 0.01~0.2) to ensure the model can both track load changes and maintain stability. For the intraday rolling forecasting model (multiple linear regression), a least mean square filtering algorithm is used to update the regression coefficient vector. Normalization is achieved through the Euclidean norm of the input vector, while a threshold is set to prevent numerical divergence. The above calculations can be deployed on a local computer or cloud platform and executed every 15 minutes.

[0038] For example, the parameters of the day-ahead cycle forecasting model and the intraday rolling forecasting model, and their specific implementation process include: Obtain the actual load value L at time t. actual,t , respectively with the predicted values ​​L of the two prediction models t,1 L t,2 Compare and calculate the predicted residual G t,1 and G t,2 : ; Using the predicted residual Gt, the parameters of the day-ahead cycle prediction model and the intraday rolling prediction model in step S2 are updated online using a recursive algorithm with a forgetting factor: For the day-ahead cycle forecasting model, its horizontal component, trend component, and seasonal component are recursively corrected: ; ; ; For the intraday rolling forecast model, the original input vector and regression coefficient vector are first standardized (Z-score). The original input vector is: ; The least mean square filtering algorithm is used to recursively update the regression coefficient vector in the intraday rolling forecast model: ; Among them, h1, h2, h3, and h4 are forgetting factors, which are used to control the parameter update speed and take values ​​of 0.01 to 0.2; H t+1 and H t These are the standardized regression coefficient vectors at times t+1 and t, respectively. ; Let t be the Euclidean norm of the standardized input vector of the intraday rolling forecast model at time t; c is the smoothing constant, which is set to 10. -6 kW, dimensionless.

[0039] According to the above technical solution, the online learning mechanism provided by this invention enables the day-ahead cycle prediction model and the intraday rolling model to continuously adapt to changes in system characteristics, such as equipment aging and seasonal changes, which can reduce the accumulation of prediction errors over time. The forgetting factor mechanism balances the influence of historical experience and the latest data, avoiding overfitting of the model to short-term fluctuations while ensuring the ability to track long-term trends. Finally, this invention also sets a threshold protection mechanism to ensure the stability of the system under abnormal operating conditions, such as sensor failure leading to data anomalies, which can effectively prevent model divergence caused by erroneous updates and significantly improve system reliability.

[0040] In one of the technical solutions, the method for obtaining the weight coefficients ω1, ω2, and ω3 of the comprehensive benefit function in step S4 includes: Based on historical meteorological data and load datasets, corresponding refrigeration unit operation records are used to construct a training sample set containing various operating conditions. Under the constraint that ω1+ω2+ω3=1 and all terms are greater than or equal to 0, generate multiple sets of candidate weight coefficient combinations; Using the training sample set as input, the prediction and energy efficiency control method of the parallel system of the refrigeration unit is simulated to obtain the simulated operation command sequence and the corresponding total cooling capacity, actual load and equipment start-up and shutdown records of the system. Based on simulation records, construct long-term comprehensive performance indices corresponding to weighting coefficients ω1, ω2, and ω3: ; The combination of weight coefficients ω1, ω2, and ω3 that minimizes the long-term comprehensive performance index can be searched using a genetic algorithm or a particle swarm optimization algorithm. Among them, E total The total energy consumption of all cooling units in the system during the training period, in kW·h; D total The integral of the absolute value of the deviation between the total cooling capacity of the system and the actual load at each moment during the training period, expressed in kW·h. S total This is the sum of the start and stop times of all refrigeration units during the training period; λ1 is D total Conversion factor, value 1~10, dimensionless; λ2 is S total The conversion factor, ranging from 1 to 50, is used to convert the number of start-stop cycles into equivalent energy consumption.

[0041] This invention further utilizes an intelligent optimization algorithm to automatically determine the weighting coefficients in the comprehensive benefit function, achieving an optimal balance between energy efficiency, load matching, and equipment switching. The system constructs a training sample set based on historical meteorological data (outdoor temperature and humidity records from the past year) and corresponding refrigeration unit operating records (including COP value, cooling capacity, start-stop frequency, etc.), covering different seasons and various load conditions. First, multiple candidate combinations of weighting coefficients are generated under the constraint ω1+ω2+ω3=1. Then, for each candidate weighting set, the refrigeration unit parallel system prediction and energy efficiency control method is fully simulated on the training sample set, and the total system energy consumption E during the simulation is recorded. total Cumulative load tracking absolute deviation D total Total number of device start-stops S total Based on these physical quantities, a long-term comprehensive performance index is constructed, with λ1 and λ2 as preset penalty coefficients used to convert deviations and start-up / shutdown costs to the same dimension as energy consumption. Subsequently, a genetic algorithm (population size 50-100, iterations 200-500) or a particle swarm optimization algorithm (particles 30-50) is used to solve the long-term comprehensive performance index. The algorithm sets a convergence threshold (e.g., objective function change less than 0.001) and a maximum number of iterations to prevent getting trapped in local optima. Finally, the combination of weight coefficients (ω1, ω2, ω3) that minimizes the long-term comprehensive performance index is selected as the optimal weight. This weight optimization algorithm can be embedded in a local server or a cloud platform.

[0042] The aforementioned technical solution, through data-driven weighted optimization, enables the system to automatically adapt to different building cooling characteristics and usage habits, such as the large day-night load differences in office buildings and the stable load of data centers, thereby improving overall energy efficiency. Multi-objective optimization balances the conflict between energy efficiency and equipment wear, effectively reducing the start-up and shutdown frequency of the chillers while ensuring efficient system operation, significantly extending the lifespan of the chillers. Most importantly, the optimization process is fully automated, lowering the system commissioning threshold and maintenance costs, making it particularly suitable for complex scenarios with multiple chillers connected in parallel.

[0043] The present invention further claims protection for the control system of the predictive and energy efficiency control method for the parallel system of the refrigeration unit, comprising: The energy efficiency mapping table construction module is used to construct a three-dimensional energy efficiency mapping table to characterize the performance of each refrigeration unit under varying operating conditions. The three-dimensional energy efficiency mapping table takes the real-time inlet temperature and outlet temperature of the cooling medium of the refrigeration unit and the load rate of the refrigeration unit as input variables, and its corresponding real-time energy efficiency value COP as output variables. The load forecasting module is used to execute the day-ahead periodic forecasting model and the intraday rolling forecasting model in parallel based on historical load data and multi-source environmental parameters, and generate day-ahead load forecasting values ​​and intraday load forecasting values ​​respectively. The multi-source environmental parameters include at least outdoor dry-bulb temperature, wet-bulb temperature and relative humidity. The credibility fusion module is used to introduce a credibility dynamic allocation mechanism based on the inverse ratio of prediction error, calculate the credibility coefficient of the day-ahead load forecast and the intraday load forecast in real time, and perform weighted fusion of the two load forecasts according to the credibility to generate a comprehensive load forecast. The credibility coefficient is dynamically adjusted according to the absolute error between the load forecast and the actual load in the previous period. The optimization control module is used to enumerate all feasible chiller operation combinations with the comprehensive load forecast value as the expected output. For each chiller operation combination, it queries the three-dimensional energy efficiency mapping table to obtain the real-time energy efficiency value of each chiller, calculates the overall energy efficiency of the combined system, constructs a comprehensive benefit function, and solves the maximum value of the comprehensive benefit function to dynamically determine the optimal number of chillers to operate and the combination strategy. It compares the current chiller operation combination with the optimal chiller operation combination, generates and issues instructions to load or unload the chillers. The model update module is used to collect actual load values, calculate the residuals between them and the day-ahead load forecast values ​​and the intraday load forecast values, and use the residuals to update the parameters of the day-ahead periodic forecast model and the intraday rolling forecast model online through a recursive algorithm, and automatically correct the day-ahead periodic forecast model and the intraday rolling forecast model for subsequent forecast periods.

[0044] The control system based on the predictive and energy efficiency control method for parallel chiller systems provided by this invention adopts a modular design, making it highly scalable and adaptable to chiller clusters of different sizes. Each module can be independently upgraded and maintained, ensuring stable system operation even under extreme conditions and preventing frequent start-ups and shutdowns or overload shutdowns of the chillers, thus reducing the actual failure rate. The entire control system can use standardized communication protocols, ensuring seamless integration with mainstream building automation systems (such as Johnson Controls Metasys and Siemens Desigo), reducing system modification costs.

[0045] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention. Applications, modifications, and variations of the present invention's method and system for predictive and energy efficiency control of parallel refrigeration systems will be readily apparent to those skilled in the art.

[0046] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for refrigeration host and associated system prediction and energy efficiency control, characterized in that, Including: S1. Construct a three-dimensional energy efficiency mapping table for each refrigeration host, where the three-dimensional energy efficiency mapping table takes the operating condition parameters and load rate of the refrigeration host as input variables and the real-time energy efficiency value COP as the output variable; S2. Based on historical load data and environmental parameters, parallelly execute the day-ahead cycle prediction model and the intraday rolling prediction model, and respectively generate the day-ahead load prediction value and the intraday load prediction value; S3. According to the errors between the day-ahead load prediction value and the intraday load prediction value in step S2 and the actual load value, dynamically calculate the credibility coefficients of the two load prediction values, and perform weighted fusion on the two load prediction values based on the credibility coefficients to generate a comprehensive load prediction value; S4. Take the comprehensive load prediction value obtained in step S3 as the expected output, enumerate all feasible refrigeration host operation combinations, for each group of refrigeration host operation combinations, query the three-dimensional energy efficiency mapping table in step S1 to obtain the real-time energy efficiency values of each refrigeration host, construct a comprehensive revenue function, determine the optimal refrigeration host operation combination with the maximization of the comprehensive revenue function as the optimization goal, and generate and issue control instructions for loading or unloading the refrigeration host; S5. Collect the actual load value, calculate its residuals with the day-ahead load prediction value and the intraday load prediction value respectively, and use the respective residuals to online update the parameters of the day-ahead cycle prediction model and the intraday rolling prediction model in step S2, and automatically correct the day-ahead cycle prediction model and the intraday rolling prediction model in subsequent prediction cycles; Among them, the comprehensive revenue function is at least related to the system average energy efficiency level, the deviation level between the total cooling capacity and the comprehensive load prediction value, and the number of start-stop times of the refrigeration host.

2. The refrigeration host and system prediction and energy efficiency control method of claim 1, wherein, The comprehensive yield function in the step S4 is: The optimization objective is: ； where F i is the overall benefit value of the ith operating combination, dimensionless; Ei = average energy efficiency normalized value for the ith operating combination, dimensionless; ; COP i COPi is the average energy efficiency of the i-th operating combination; COP ref Take the standard COP value of the refrigeration host under rated working condition; Q i Qtotal,i total refrigeration capacity for the ith operating combination, kW; L pred To integrate the load forecast value, kW; Normalized value of the deviation of the total refrigeration capacity and the combined load forecast value for the i-th operating combination, dimensionless; ; Q ref Qnorm is the normalized reference value, taken as the rated total refrigeration capacity of the system, in kW. a normalized value of the number of start / stop of the refrigeration host required to switch from the current state to the i-th combination; N is the number of times of switching from the current state to the refrigeration host start-stop in the i-th combination; N ref Take the preset single maximum allowed start-stop number; ω1, ω2, ω3 are weight coefficients for each item, dimensionless, obtained through an optimization algorithm of historical operation data, and ω1 + ω2 + ω3 = 1.

3. The refrigeration host and system prediction and energy efficiency control method of claim 2, wherein, The day-ahead cycle load model in step S2 adopts the Holt-Winters additive model, which is used to predict the load profile for the next 24-hour cycle, specifically including: Based on historical load time series data, initialize the level component, trend component, and seasonal component of the Holt-Winters additive model; For each moment t, use the following recurrence formula to dynamically update the level component, trend component, and seasonal component of the Holt-Winters additive model according to the latest monitored load value: Horizontal component: ; Trend component: ; Seasonal component: ; Based on the updated level component, trend component, and seasonal component of the Holt-Winters additive model, generate the load prediction value for the next k steps as the day-ahead load prediction value: Load prediction value for future k steps ; wherein l t , l t-1 is the model horizontal component, respectively the load estimate value at the current time t and at the previous time t-1, kW; y t Pact is the actual load measurement value for the current time t, kW; b t , b t-1 is the model trend component, respectively the load change amount at the current time t and at the previous time t-1, kW; s t t-p Seasonal component of the model, seasonal load deviation value at current time t and at time t one period p ago, respectively, kW​ p is the cycle length, 24h; γ, ε, and μ are respectively the level smoothing coefficient, trend smoothing coefficient, and seasonal smoothing coefficient. γ takes a value of 0.1 - 0.3, dimensionless; ε takes a value of 0.01 - 0.2, dimensionless; μ takes a value of 0.2 - 0.5, dimensionless; k is the prediction step length, k < p, dimensionless; P(t+k) is the load forecast value for time t+k, kW; s t-p+k Seasonal load deviation value, kW, for one period p back in time for time instant t+k.

4. The refrigeration host and system prediction and energy efficiency control method of claim 3, wherein, The intraday rolling prediction model in step S2 is constructed based on multiple linear regression and is used to achieve short-term and fine load prediction in the future, specifically including: Based on the latest monitored historical load data, historical prediction residuals, and real-time meteorological parameters, construct the intraday rolling prediction model: ； where L i,t is the intraday load forecast value output by the intraday rolling prediction model at the current time t, kW; θ0 is the constant of the intraday rolling prediction model, kW; L t-j Lj(t) is the measured historical load value for the jth hour before the current time t, kW; θ j Regression coefficient for historical load term, dimensionless; m represents the number of historical load items considered in the intraday rolling forecast model, which is dimensionless. B t-q is the difference between the historical forecast value and the measured value for the time instant t - q hours, kW; A q Regression coefficient for historical prediction residual term, dimensionless; n is the number of historical forecast residual items considered in the intraday rolling forecast model, which is dimensionless; D1 is the regression coefficient of the outdoor temperature term, kW / ℃; T out Tout is the outdoor dry-bulb temperature monitored at the current time t, °C; D2 is the regression coefficient for the outdoor relative humidity term, kW / %; RH represents the outdoor relative humidity monitored at time t, in %.

5. The refrigeration host and system prediction and energy efficiency control method of claim 4, wherein, The specific implementation process of weighted fusion of the two load forecast values ​​to generate the comprehensive load forecast value in step S3 includes: Based on the latest forecast errors of the day-ahead cycle forecast model and the intraday rolling forecast model, the reliability coefficients of the day-ahead load forecast and the intraday load forecast are calculated using the following method: ； R2 = 1 - R1; The obtained day-ahead load prediction value and the credibility coefficient of the intra-day load prediction value are used to weight and fuse the two load prediction values to generate a comprehensive load prediction value L pred : ； Wherein, R1 is the confidence coefficient of the day-ahead load forecast, with a value of 0 to 1; R2 is the confidence coefficient of the intraday load forecast, with a value of 0 to 1; L1 is the daytime load forecast value obtained in step S2, in kW; L2 is the daily load forecast value obtained in step S2, in kW; L actual Actual Load Measurement, kW; c is a smoothing constant, taken as 10 -6 kW; L pred To synthesize the load forecast value, kW.

6. The refrigeration host and system prediction and energy efficiency control method of claim 5, wherein, The methods for obtaining the weight coefficients ω1, ω2, and ω3 of the comprehensive benefit function in step S4 include: Based on historical meteorological data and load datasets, corresponding refrigeration unit operation records are used to construct a training sample set containing various operating conditions. Under the constraint that ω1+ω2+ω3=1 and all terms are greater than or equal to 0, generate multiple sets of candidate weight coefficient combinations; Using the training sample set as input, the prediction and energy efficiency control method of the parallel system of the refrigeration unit is simulated to obtain the simulated operation command sequence and the corresponding total cooling capacity, actual load and equipment start-up and shutdown records of the system. Based on simulation records, construct long-term comprehensive performance indices corresponding to weighting coefficients ω1, ω2, and ω3: ； The combination of weight coefficients ω1, ω2, and ω3 that minimizes the long-term comprehensive performance index can be searched using a genetic algorithm or a particle swarm optimization algorithm. wherein E total is the total energy consumption of all refrigeration hosts in the system during the training period, kW·h; D total is the absolute value of the deviation between the total refrigeration capacity of the system and the actual load at each time during the training period, kW·h. S total For the training period, the sum of the start-stop times of all refrigeration hosts is next. λ1 is D total conversion factor, value 1-10, dimensionless; λ2 is S total Conversion factor, 1~50; unit is kW·h / time.

7. A control system based on the refrigeration host machine of any one of claims 1 to 6 in conjunction with the system prediction and energy efficiency control method, characterized in that, include: The energy efficiency mapping table construction module is used to construct a three-dimensional energy efficiency mapping table to characterize the performance of each refrigeration unit under varying operating conditions. The three-dimensional energy efficiency mapping table takes the real-time inlet temperature and outlet temperature of the cooling medium of the refrigeration unit and the load rate of the refrigeration unit as input variables, and its corresponding real-time energy efficiency value COP as output variables. The load forecasting module is used to execute the day-ahead periodic forecasting model and the intraday rolling forecasting model in parallel based on historical load data and multi-source environmental parameters, and generate day-ahead load forecasting values ​​and intraday load forecasting values ​​respectively. The multi-source environmental parameters include at least outdoor dry-bulb temperature, wet-bulb temperature and relative humidity. The credibility fusion module is used to introduce a credibility dynamic allocation mechanism based on the inverse ratio of prediction error, calculate the credibility coefficient of the day-ahead load forecast and the intraday load forecast in real time, and perform weighted fusion of the two load forecasts according to the credibility to generate a comprehensive load forecast. The credibility coefficient is dynamically adjusted according to the absolute error between the load forecast and the actual load in the previous period. The optimization control module is used to enumerate all feasible chiller operation combinations with the comprehensive load forecast value as the expected output. For each chiller operation combination, it queries the three-dimensional energy efficiency mapping table to obtain the real-time energy efficiency value of each chiller, calculates the overall energy efficiency of the combined system, constructs a comprehensive benefit function, and solves the maximum value of the comprehensive benefit function to dynamically determine the optimal number of chillers to operate and the combination strategy. It compares the current chiller operation combination with the optimal chiller operation combination, generates and issues instructions to load or unload the chillers. The model update module is used to collect actual load values, calculate the residuals between them and the day-ahead load forecast values ​​and the intraday load forecast values, and use the residuals to update the parameters of the day-ahead periodic forecast model and the intraday rolling forecast model online through a recursive algorithm, and automatically correct the day-ahead periodic forecast model and the intraday rolling forecast model for subsequent forecast periods.

Images