Global optimization method for air conditioning system and storage medium

Abstract

The application provides a global optimization method of an air conditioning system and a storage medium, and relates to the technical field of intelligent control of air conditioning systems. The application first screens target variables through global sensitivity analysis and generates high-quality training data. Then, the grid search is used to optimize the hyperparameters and verify the generalization ability. Then, the multi-class global optimization algorithm is used to optimize the optimization variables and extract performance data. Based on multi-dimensional trade-off analysis, the target combination of the model and the algorithm is determined to achieve the collaborative optimization of the prediction accuracy, optimization accuracy and optimization speed. Finally, the target combination is embedded in the global optimization architecture and combined with the interference factor prediction value to output the set value, ensuring that the optimization scheme can dynamically adapt to environmental and load changes, significantly reducing the total energy consumption of the air conditioning system, while ensuring the system operation stability and indoor thermal comfort, and realizing the global optimal operation effect.

Description

Global optimization methods and storage media for air conditioning systems Technical Field

[0001] This application relates to the field of intelligent control technology for air conditioning systems, and more specifically, to a global optimization method and storage medium for an air conditioning system. Background Technology

[0002] In the field of intelligent control of traditional air conditioning systems, control methods often have many limitations. On the one hand, most methods lack in-depth analysis of data characteristics when collecting air conditioning system operating data (such as chilled water inlet / outlet temperature and flow rate) and environmental data (such as temperature and humidity). Specifically, during the operation of an air conditioning system, various actuators generate a large number of real-time operating status characteristics, while indoor and outdoor environments also bring various environmental impact characteristics. However, existing technologies usually simply summarize and process the above data without identifying and removing interference characteristics that have low correlation with changes in air conditioning system load and total energy consumption, resulting in data redundancy and high noise interference. Due to the strong coupling, nonlinearity, and long time delay characteristics of air conditioning systems, coupled with the influence of interference data, traditional optimization methods struggle to predict their complex physical behavior. When the behavior is non-convex, conventional optimization algorithms rarely obtain the global optimal solution, ultimately causing significant deviations in the prediction results of core target parameters such as total energy consumption.

[0003] On the other hand, existing air conditioning system control methods fail to effectively construct a cross-domain coupling mechanism between the operating characteristics of the air conditioning system and the characteristics of environmental impact when establishing data relationships. Specifically, there are complex dynamic correlations between different characteristics, and these dynamic correlations evolve dynamically with time and operating conditions. Traditional methods often use static modeling methods, which are difficult to capture the above-mentioned dynamic coupling relationships and cannot accurately reflect the operating patterns of the air conditioning system under different environments and load conditions.

[0004] Furthermore, in the load forecasting stage, existing technologies often lack effective verification and calibration mechanisms for forecast results. Relying solely on the output of a single forecasting model without incorporating historical forecasts and deviations from actual loads to correct current forecasts leads to significant discrepancies between forecasts and actual loads, making it difficult to meet the requirements for precise operation and control of air conditioning systems. In generating operational control sequences, there is also a lack of comprehensive consideration of multiple objectives and priorities, making it difficult for air conditioning systems to simultaneously address multiple important objectives such as load matching, energy consumption optimization, and operational stability during operation. This results in frequent problems such as insufficient energy saving, load mismatch, or large operational fluctuations, hindering the achievement of globally optimal coordination. Summary of the Invention

[0005] The purpose of this application is to provide a global optimization method and storage medium for an air conditioning system, in order to solve the problem that traditional air conditioning systems are difficult to balance load matching, energy consumption optimization and operational stability during operation.

[0006] To achieve the above objectives, the first aspect of this application provides a global optimization method for an air conditioning system, the method comprising:

[0007] Historical operating data of the air conditioning system and corresponding indoor and outdoor environmental data are collected. Global sensitivity analysis is performed using the Morris screening method to determine the target variable affecting the total energy consumption of the air conditioning system. Training data is obtained based on the target variable. The global sensitivity analysis is used to quantify the main effects of single variables and the coupling effects between variables.

[0008] The training data is input into the training architecture of the pre-constructed multi-class initial total energy consumption prediction model of the air conditioning system. The optimal hyperparameters of each class of the initial total energy consumption prediction model are determined by the differentiated grid search method, and the generalization ability of the initial total energy consumption prediction model is verified to obtain the optimized target total energy consumption prediction model.

[0009] Based on the preset range of values ​​for optimization variables and in association with the indoor thermal comfort requirements of the air conditioning system, a two-layer constraint system with a penalty term is constructed. The optimization variables in the target total energy consumption prediction model are optimized using a global optimization algorithm to obtain optimization performance data corresponding to different global optimization algorithms.

[0010] Based on the optimized performance data, through a multi-dimensional trade-off analysis of prediction accuracy, optimization accuracy, and optimization speed, the objective weights are calculated using the entropy weight method, and a correction weight for optimization accuracy under high load conditions is assigned to determine the target combination of the target total energy consumption prediction model and the global optimization algorithm.

[0011] By combining the predicted values ​​of interference factors, the target combination is embedded into the global optimization architecture of the air conditioning system. After prediction modeling and global optimization corresponding to the target combination, the target set value of the air conditioning system is output.

[0012] In another aspect, this application also provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores a program that can be loaded by a processor and executed by the above-described global optimization method for an air conditioning system.

[0013] Based on the above, target variables are screened and high-quality training data is generated through global sensitivity analysis to accurately identify the core factors affecting total energy consumption, laying a reliable data foundation for subsequent model training and reducing interference from redundant variables. Then, grid search is used to optimize hyperparameters and verify generalization ability, enabling the target total energy consumption prediction model to accurately fit the energy consumption pattern of the air conditioning system, effectively reducing prediction errors and providing accurate basis for optimization decisions. Multiple global optimization algorithms are used to optimize variables and extract performance data, comprehensively exploring the system's energy-saving potential and providing rich reference dimensions for subsequent combination selection. Based on multi-dimensional trade-off analysis, the target combination of model and algorithm is determined to achieve synergistic optimization of prediction accuracy, optimization accuracy, and optimization speed, reducing the optimization limitations caused by a single indicator. The target combination is embedded into the global optimization architecture and combined with the predicted values ​​of interference factors to output setpoints, ensuring that the optimization scheme can dynamically adapt to environmental and load changes. While significantly reducing the total energy consumption of the air conditioning system, it also ensures system operational stability and indoor thermal comfort, achieving globally optimal operating results. Attached Figure Description

[0014] Figure 1 is a schematic diagram of the execution flow of a global optimization method for an air conditioning system provided in an embodiment of this application.

[0015] Figure 2 is a schematic diagram of a global optimization architecture for an air conditioning system provided in an embodiment of this application.

[0016] Figure 3 is a schematic diagram of a rolling optimization process provided in an embodiment of this application.

[0017] Figure 4 is a schematic diagram of a non-convex surface of an objective function provided in an embodiment of this application.

[0018] Figure 5 is a schematic diagram of a global optimization architecture for an air conditioning system provided in a specific embodiment of this application.

[0019] Figure 6 is a schematic diagram illustrating the application of a global optimization method for an air conditioning system in an embodiment of this application. Detailed Implementation

[0020] The present application will now be described in detail with reference to the accompanying drawings. Figure 1 is a flowchart illustrating a global optimization method for an air conditioning system according to an embodiment of the present application. The global optimization method for the air conditioning system will now be described in detail.

[0021] Step 101: Collect historical operating data of the air conditioning system and corresponding indoor and outdoor environmental data. Perform global sensitivity analysis using the Morris screening method to determine the target variable affecting the total energy consumption of the air conditioning system, and obtain training data based on the target variable. Global sensitivity analysis is used to quantify the main effects of single variables and the coupling effects between variables.

[0022] Historical operating data consists of core operating parameters recorded during the past operation of the air conditioning system. It reflects the actual working status of the system equipment and serves as the basic data source for modeling. Indoor and outdoor environmental data are environmental parameters that affect the operation of the air conditioning system, including outdoor wet-bulb temperature (reflecting outdoor meteorological conditions) and indoor cooling load (reflecting indoor cooling demand).

[0023] In this embodiment, the target variable is a core variable that significantly affects the total energy consumption of the air conditioning system, selected through global sensitivity analysis. The target variable may include temperature data, flow rate data, and load and power data. Temperature data may include chilled water outlet temperature, cooling water inlet temperature, and outdoor wet-bulb temperature. Flow rate data may include chilled water mass flow rate and cooling water mass flow rate. Load and power data may include cooling load, outdoor wet-bulb temperature, and total power of the storage medium. After collecting historical operating data and indoor / outdoor environmental data, preprocessing can be performed on the historical operating data and indoor / outdoor environmental data.

[0024] Specifically, the 3σ criterion is used to identify and remove outliers for chilled water outlet temperature, cooling water inlet temperature, and outdoor wet-bulb temperature. In one example, the data can be segmented according to time sequence, dividing the three types of temperature data into several independent time-segment datasets to reduce interference from natural temperature fluctuations across time periods and outlier identification. For each time-segment dataset, the mean and standard deviation are calculated. Then, a dynamic threshold is set based on the 3σ criterion, where the value of 3σ is adapted to the normal fluctuation characteristics of air conditioning temperature data, reducing the misjudgment of normal temperatures caused by extreme weather as outliers. Then, combined with the physical limits of air conditioning equipment operation, an absolute threshold is set. If the data exceeds the physical boundary, it is judged as an outlier regardless of whether the 3σ criterion is met. For the initially identified suspected outliers exceeding [μ - 3σ, μ + 3σ] or physical boundaries, the correlation between adjacent data is verified. For example, check the variation range of three consecutive data points before and after the given data point. If the adjacent data points are all within the normal range and the variation is stable (difference between adjacent data points ≤ 0.5℃), then the data is confirmed to be an outlier caused by instantaneous fluctuations in the sensor and should be removed. For the missing positions after removing outliers, linear interpolation is used to fill in the missing data to ensure the continuity of the data sequence.

[0025] For chilled water mass flow rate and cooling water mass flow rate, the original cumulative flow rate data is converted into unit-time flow rate data based on the sampling time interval and the original cumulative flow rate data. In one example, valid data segments with monotonically increasing cumulative values ​​in the original cumulative flow rate data can be filtered out, and data with decreasing or abrupt changes in cumulative values ​​due to sensor malfunctions can be removed. The data sampling time interval is confirmed to be consistent with the preset interval for air conditioning system data acquisition. If there are individual sampling interval deviations, the sampling time interval is calculated based on the actual timestamp. For valid cumulative flow rate data, the unit-time flow rate is calculated according to continuous time sequence. For the first data point, the average unit-time flow rate of the subsequent three data points is used to back-complete the data, reducing the loss of initial data. Based on the rated flow range of the air conditioning water pump, outliers where the converted unit-time flow rate exceeds the rated range are identified. The coefficient of variation of the converted flow rate data is calculated. If the coefficient of variation is greater than the set stability threshold, the original cumulative flow rate data is retrieved, the data validity is re-verified, and the conversion steps are repeated.

[0026] For cooling load and total system power, dimensional normalization is performed, mapping the values ​​to the [0,1] interval. In one example, based on the cooling load and total system power data after outlier identification, valid data within the set rated operating range are further filtered out. The maximum and minimum values ​​of the two types of data are recorded, and extreme value parameters are saved. Then, the Min-Max normalization method is used to map each data point to the [0,1] interval. For example, based on the formula y_norm = (y_j -Min) / (Max - Min), if Max = Min (e.g., the system operates at a low load for a long time, resulting in no data fluctuation), then this type of data is uniformly mapped to 0.5 (reducing calculation errors due to a denominator of 0, while ensuring the effectiveness of data participation in model training). The normalized data undergoes interval verification. If there are outliers such as y_norm < 0 or y_norm > 1, truncation is performed (values ​​less than 0 are set to 0, and values ​​greater than 1 are set to 1) to ensure that all data fall within the [0,1] interval.

[0027] Global sensitivity analysis is an analytical method that quantifies the influence of each input variable on the output response (such as total energy consumption). It is used to screen key influencing variables and reduce interference from redundant information. In this application, the Morris screening method can be used to determine the target variable affecting the total energy consumption of the air conditioning system, and training data can be obtained based on the target variable. The training data is a dataset used for model training and validation, obtained after preprocessing and partitioning following the selection of the target variable, ensuring data quality and suitability.

[0028] Specifically, using total system power as the response indicator for total energy consumption, a set of candidate variables strongly correlated with the operation of the air conditioning system is selected. For example, based on the physical relationship "total energy consumption = ∫ total system power dt (time integral)," the dynamic characteristics of total energy consumption are indirectly characterized through real-time power data, ensuring both convenient data acquisition and accurate reflection of the correlation between energy consumption and variables. Time-series smoothing is applied to the total system power data to eliminate the interference of instantaneous power fluctuations on the response indicator, ensuring indicator stability. The candidate variable set consists of variables strongly correlated with the total energy consumption of the air conditioning system, covering three dimensions: equipment operating parameters, environmental influencing factors, and load demand. As an example, the candidate variable set may include the aforementioned target variables. Chilled water outlet temperature and chilled water mass flow rate directly affect the chiller unit's coefficient of performance and evaporator heat exchange efficiency; cooling water system parameters are related to the condenser's heat dissipation effect, affecting unit energy consumption. Indoor cooling load determines the cooling output demand of the air conditioning system. Outdoor wet-bulb temperature affects the cooling tower's cooling effect, indirectly affecting the cooling water temperature. The total system power is used as a common parameter for the response index to help quantify the autocorrelation characteristics. The Pearson correlation coefficient between each variable and the total system power is calculated. Variables with an absolute value of the correlation coefficient greater than a set coefficient are retained, while variables with weak correlation are removed, ultimately forming a candidate variable set.

[0029] Then, the number of variable levels adapted to the multivariate coupling characteristics of air conditioning and the number of estimated levels covering typical operating conditions of air conditioning are set, and the sampling trajectory of the variables is constructed. For example, considering the characteristics of strong multivariate coupling and continuous parameter changes in air conditioning systems, each candidate variable can be set with 4-5 levels, and the level division can adopt the principles of equidistant and operating condition adaptation. For example, temperature variables (chilled water outlet temperature, cooling water inlet temperature, outdoor wet-bulb temperature): three basic levels are defined at equal intervals according to the rated operating range of the equipment. Two additional densified levels are added in the parameter range corresponding to high-load conditions (e.g., chilled water outlet temperature 5-12℃, basic levels 5 / 9 / 12℃, high-load densified levels 7 / 10℃). Flow variables (chilled water mass flow rate, cooling water mass flow rate): five levels are set according to 20% / 40% / 60% / 80% / 100% of the pump's rated flow rate range, covering the full range of flow rate adjustment. Cooling load variables: four levels are set according to 30% / 50% / 70% / 90% of the system's rated cooling capacity, matching typical low / medium / full load conditions. The estimated number of typical air conditioning operating conditions (i.e., the number of sampling trajectories in the Morris screening method) is set to 15-20, determined based on the distribution of typical operating conditions of the air conditioning system. The historical operating data is analyzed to determine the proportion of high, medium, and low load conditions (e.g., 30% high load, 50% medium load, 20% low load). The number of sampling trajectories corresponding to each condition is allocated according to this proportion (e.g., 5 for high load, 10 for medium load, and 5 for low load). Each trajectory is ensured to cover parameter combinations of at least two different operating conditions to avoid analytical bias caused by trajectories focusing on a single condition. Based on the Morris method's "One Variable at a Time (OAT)" principle, the sampling order is adjusted according to the coupling relationship of air conditioning variables: for highly coupled variable pairs (e.g., chilled water outlet temperature and cooling load, cooling water inlet temperature and outdoor wet-bulb temperature), these pairs are preferentially alternated in the sampling trajectories to enhance the capture of coupling effects; the parameter combinations of the sampling trajectories must meet equipment operating constraints (e.g., chilled water outlet temperature not lower than 5℃), and invalid combinations exceeding physical boundaries are eliminated; finally, a three-dimensional adaptive sampling trajectory matrix is ​​formed, consisting of "variable level × operating condition distribution × coupling relationship," with the matrix dimension being "number of trajectories × number of candidate variables."

[0030] Next, the global sensitivity index for each candidate variable in the candidate variable set is calculated. The global sensitivity index is used to quantify the impact of a single variable and the coupling effect between variables on total energy consumption. The candidate variable data in the sampling trajectory are aligned with the corresponding system total power response index data to form a one-to-one correspondence dataset of "variable combination - response value". Normalization processing is performed on all variable data (mapped to the [0,1] interval) to eliminate the influence of dimensional differences on index calculation and ensure the comparability of sensitivity of different types of variables. The global sensitivity index calculation model of Morris screening method is adopted to quantify the main effect of a single variable and the coupling effect between variables. For example, the main effect index is calculated as the mean change of response value of each variable at different levels, and the coupling effect index is calculated as the standard deviation of the change of response value of each variable, reflecting the strength of the coupling effect between the variable and other variables. The comprehensive global sensitivity index is a weighted sum that integrates the main effect and the coupling effect. The main effect index has a slightly higher weight to adapt to the energy consumption characteristics of the air conditioning system, which is "main effect dominant and coupling effect auxiliary". The global sensitivity index is calculated three times and the mean is taken as the final result to reduce the error caused by sampling randomness. If the overall sensitivity index S of a variable is greater than or equal to 0.1, and both the main effect index and the coupling effect index are not 0, then the index calculation of the variable is considered valid. If there is a variable with a coupling effect index of 0, it means that it has no obvious coupling with other variables, and only the main effect index is retained as the basis for sensitivity evaluation.

[0031] Next, based on the historical total energy consumption fluctuation confidence interval of the air conditioning system, a sensitivity index threshold is dynamically set, and candidate variables with a global sensitivity index greater than or equal to the threshold are selected as target variables. In one example, historical total energy consumption data of the air conditioning system (daily cumulative energy consumption) can be extracted, and a 95% confidence interval is calculated using a normal distribution fitting method. The sensitivity index threshold is dynamically adjusted based on the fluctuation amplitude of the energy consumption fluctuation confidence interval, where the formula can satisfy: T=T0×(1+ΔE / μE), where T0 is the baseline threshold (taken as 0.2, calibrated based on the empirical value of similar optimization methods in the industry). Candidate variables with a comprehensive global sensitivity index greater than or equal to the dynamic threshold T are selected as core target variables affecting total energy consumption. Coupled coverage verification is performed on the selected target variables: ensuring that variables are included in all four dimensions—chilled water system, cooling water system, load, and environment—to avoid one-sided model input caused by missing dimensions. If no variable is selected for a certain dimension (such as outdoor wet-bulb temperature S < T in the environmental dimension), the sensitivity index threshold of the variable for that dimension is lowered by 10%, and the selection is repeated to ensure that each dimension is covered by core variables, thus forming the target variable set.

[0032] Finally, based on the temporal continuity of historical operational data, the target variable is divided into training and validation sets according to the time interval ratio. For example, the training and validation sets are divided in an 8:2 time interval ratio. If the total duration of historical data is 120 days (a complete cooling season), then 96 days (80%) of the data are used as the training set, and the remaining 24 days (20%) are used as the validation set. The validation set must cover all typical operating conditions: ensure that the proportion of high / medium / low load conditions in the validation set is consistent with that in the training set (deviation ≤ 5%), and reduce the distortion of generalization ability assessment caused by the validation set being concentrated on a single operating condition. Calculate the mean and standard deviation of the target variable for both datasets, requiring that the mean deviation of each variable be ≤ 10% and the standard deviation deviation be ≤ 15%. If the statistical deviation of a variable exceeds the threshold (e.g., the mean temperature of chilled water outlet in the training set is 7℃, and in the validation set it is 9℃), then the division boundary is finely adjusted in 7-day increments (e.g., extending the training set by 7 days and shortening the validation set by 7 days), and re-verification is performed until the consistency requirements are met. The final output consists of a training set and a validation set that are time-continuous and have a consistent distribution, which are used for training and generalization capability verification of the total energy consumption prediction model.

[0033] Step 102: Input the training data into the training architecture of the pre-built multi-class initial total energy consumption prediction model of the air conditioning system, use the differentiated grid search method to determine the optimal hyperparameters of each class of initial total energy consumption prediction model, and verify the generalization ability of the initial total energy consumption prediction model to obtain the optimized target total energy consumption prediction model.

[0034] The initial total energy consumption prediction model refers to a pre-built set of basic machine learning models that have not undergone hyperparameter optimization. It encompasses both linear and nonlinear model types and is used to initially fit the mapping relationship between the total energy consumption of an air conditioning system and various influencing variables. The training data corresponding to the target variable is standardized and adapted according to the input format requirements of the initial total energy consumption prediction model to ensure that the data can be directly input into the model for training.

[0035] In one example, a set of multiple types of machine learning models can be pre-built. This set of models can include multiple initial sub-models, such as Multiple Linear Regression (MLR) models, Artificial Neural Network (ANN) models, Support Vector Regression (SVR) models, and Random Forest (RF) models, to obtain an initial total energy consumption prediction model. These multiple initial sub-models cover both linear and nonlinear modeling capabilities, adapting to the energy consumption characteristics of air conditioning systems, which exhibit a combination of linear main effects and nonlinear coupling effects.

[0036] The grid search method discretizes the range of hyperparameter values ​​into grid points, traverses all grid point combinations, evaluates model performance, and finally selects the optimal hyperparameter combination. It features clear optimization logic and stable results. The optimal hyperparameters refer to the hyperparameter combination in the grid search hyperparameter space that minimizes the mean squared error (MSE) of the initial total energy consumption prediction model on the training set. These are key parameters for adapting the model to the energy consumption characteristics of the air conditioning system.

[0037] Differentiated grid search hyperparameter spaces can be designed for different types of initial sub-models. For example, for the Multiple Linear Regression (MLR) model, the hyperparameter space is set as the regularization coefficient range. For the Artificial Neural Network (ANN) model, the hyperparameter space is set as the range of hidden layer node count and learning rate. For the Support Vector Regression (SVR) model, the hyperparameter space is set as the kernel function type and kernel function parameter range. For the Random Forest (RF) model, the hyperparameter space is set as the range of decision tree count and maximum depth. Setting the regularization coefficient range for the MLR model adapts to slight collinearity in the data. Setting the range of hidden layer node count (10-50) and learning rate (0.001-0.01) for the ANN model adapts to time-varying energy consumption characteristics. Setting the kernel function type and corresponding kernel parameter range for the SVR model adapts to nonlinear coupling characteristics; setting the range of decision tree count (50-200) and maximum depth (5-20) for the RF model adapts to multivariate interaction characteristics. Then, using the mean square error of total energy consumption prediction in the training set as the objective function, a grid search is performed simultaneously on multiple initial sub-models, traversing the hyperparameter space of each initial sub-model, and selecting the hyperparameter combination with the smallest mean square error of total energy consumption prediction as the target hyperparameter of the initial sub-model.

[0038] Next, the validation set was divided into multiple sub-validation sets according to the operating conditions of the air conditioning system. Hierarchical cross-validation was performed on each initial sub-model equipped with the target hyperparameters, and the mean square error variation coefficient of total energy consumption prediction for each initial sub-model in each sub-validation set was calculated. The operating conditions included high load, medium load, and low load conditions.

[0039] Generalization ability refers to a model's predictive power on new data (validation set data) that was not used in training. It reflects the model's ability to capture the general energy consumption patterns of air conditioning systems and reduces the likelihood of "overfitting" (fitting only the training data and not the actual operating conditions). A hierarchical cross-validation method is used, stratifying the validation set according to air conditioning operating conditions (high load, medium load, low load) to ensure that each stratum contains data from different operating conditions. For each initial model equipped with optimal hyperparameters, hierarchical cross-validation is performed, and the MSE coefficient of variation for each layer of the validation set is calculated. Only when the coefficient of variation of all layers is ≤15% is the model's generalization ability considered satisfactory, reducing the possibility of the model failing due to fitting only a single operating condition.

[0040] If the coefficient of variation of the total energy consumption prediction mean square error of the initial sub-model is less than or equal to the set coefficient of variation in all sub-validation sets, then the generalization ability of the initial sub-model is deemed qualified.

[0041] Among the initial sub-models with good generalization ability, the initial sub-model with the smallest mean square error of total energy consumption prediction in the training set and the smallest prediction error in the sub-validation set under high load conditions is selected as the target sub-model.

[0042] The initial sub-model is optimized based on the target sub-model to obtain the target total energy consumption prediction model.

[0043] The target total energy consumption prediction model, determined after hyperparameter optimization and generalization capability verification, is the final model (or model fusion) capable of accurately predicting the total energy consumption of the air conditioning system and serves as the core basis for subsequent global optimization. From the models with qualified generalization capabilities, the model with the smallest training set MSE and the smallest verification error under high load conditions is selected as the target total energy consumption prediction model. If multiple qualified models exist, a weighted fusion method is used to integrate them, with the core model having a weight of 0.6 and other qualified models having a weight of 0.4, further improving prediction reliability.

[0044] By employing multiple initial sub-models to cover different energy consumption characteristics and combining differentiated hyperparameter grid search, the model can accurately fit the linear and nonlinear energy consumption patterns of the air conditioning system, significantly reducing the average absolute percentage error of total energy consumption prediction and providing reliable data support for subsequent optimization. Hierarchical cross-validation verifies the generalization ability under different operating conditions of the air conditioning system, ensuring stable predictions under complex conditions such as high and low loads, reducing the failure of practical applications caused by fitting only a single operating condition. Determining the optimal hyperparameters solves the problem of "blindly setting parameters" in the initial model, greatly improving the model's convergence speed and prediction stability, and reducing computational redundancy in subsequent optimization processes. The high accuracy and strong generalization ability of the target total energy consumption prediction model provide a reliable objective function for the global optimization algorithm, ensuring that the algorithm can accurately tap the system's energy-saving potential, reducing optimization direction deviations caused by prediction errors, and laying a core foundation for achieving global energy optimization.

[0045] Step 103: Based on the preset range of values ​​for the optimization variables and in association with the indoor thermal comfort requirements of the air conditioning system, construct a two-layer constraint system with a penalty term. Use a global optimization algorithm to perform optimization operations on the optimization variables in the target total energy consumption prediction model to obtain optimization performance data corresponding to different global optimization algorithms.

[0046] Optimization variables refer to core operating parameters that can be directly adjusted through equipment control and have a significant impact on the total energy consumption of the air conditioning system. For example, these could be chilled water outlet temperature (°C), chilled water mass flow rate (kg / s), cooling water inlet temperature (°C), and cooling water mass flow rate (kg / s), which are the direct targets of the optimization operation. Global optimization algorithms are intelligent algorithms used to solve for the global optimum in complex nonlinear, multi-constraint spaces. This step can select three mainstream algorithms: genetic algorithm, particle swarm optimization algorithm, and simulated annealing algorithm, each focusing on its advantages in global exploration, equilibrium convergence, and avoiding local optima, respectively. The target total energy consumption prediction model is a model (or model fusion) that accurately fits the energy consumption pattern of the air conditioning system after hyperparameter optimization and generalization capability verification. In this step, it serves as an "energy consumption response calculator," providing real-time energy consumption prediction values ​​for algorithm optimization. Through multiple global optimization algorithms, the optimal values ​​of optimization variables are explored within the compliant range, while the performance differences of different algorithms are quantified, providing data support for subsequent model-algorithm combination selection. By exploring the maximum energy-saving potential of the air conditioning system and outputting algorithm performance comparison data, global optimization of optimization variables and quantification of algorithm performance are achieved. Specifically, the process includes the following steps.

[0047] First, based on the rated operating parameters of the air conditioning equipment, a safe range of values ​​for the optimization variables is set. This is then linked to the indoor thermal comfort requirements of the air conditioning system, and constraints on indoor temperature fluctuations are added, resulting in a two-layer constraint system. For example, using chilled water outlet temperature, chilled water mass flow rate, cooling water inlet temperature, and cooling water mass flow rate as core optimization variables, hard constraints (i.e., safe ranges) are set based on the equipment's rated parameters. For instance, chilled water outlet temperature: 5-12℃ (to avoid unit icing or overload); chilled water mass flow rate: 200-800 kg / s (matching the rated flow range of the chilled water pump); cooling water inlet temperature: 22-35℃ (adapting to the upper limit of cooling tower cooling capacity); cooling water mass flow rate: 300-1000 kg / s (matching the rated flow range of the cooling water pump). Soft constraints on indoor temperature fluctuations are then added in conjunction with indoor thermal comfort requirements, for example, indoor temperature fluctuation ≤ ±0.5℃. Define the logic of soft constraints: Establish a mapping relationship between optimization variables and indoor temperature using the heat exchange mechanism of the air conditioning system (e.g., a 1°C decrease in chilled water outlet temperature results in approximately a 0.3°C decrease in indoor temperature), ensuring that constraints are quantifiable and verifiable. Then, cross-validate the compatibility of hard and soft constraints: If an optimization variable satisfies a hard constraint but violates a soft constraint (e.g., a 5°C decrease in chilled water outlet temperature leads to a 0.8°C fluctuation in indoor temperature), the soft constraint is converted into a temporary hard constraint (adjusting the lower limit of the chilled water outlet temperature to 6°C). Create a two-layer constraint system document: "Hard constraints ensure equipment safety + soft constraints ensure thermal comfort," clearly defining constraint priorities (e.g., hard constraints > soft constraints). In this way, a "safety + comfort" two-layer constraint system can be constructed.

[0048] Then, differentiated optimization strategies corresponding to various global optimization algorithms are designed. The global optimization algorithms include Genetic Algorithm (GA), Particle Swarm Optimization (PSO) algorithm, and Simulated Annealing (SA) algorithm.

[0049] In one example, for the genetic algorithm, a predetermined percentage of the best individuals from the first generation are retained. The crossover probability is dynamically adjusted based on the population fitness variance. The mutation probability decreases from a first probability to a second probability with each iteration, where the first probability is greater than the second probability. For example, a real-number encoding method is used, with a population size of 50 (balancing optimization efficiency and diversity). The encoding length corresponds to four optimization variables, and each variable's encoding precision is retained to two decimal places (adapting to engineering control precision). Through an elite retention strategy, the top 10% of the best individuals from each generation are directly introduced into the next generation. The crossover probability is dynamically adjusted based on the population fitness variance. For example, when the variance ≥ 0.3 (high population diversity), the crossover probability is 0.8-0.9 (strengthening global exploration); when the variance < 0.3 (population convergence), the crossover probability is 0.5-0.7 (strengthening local development). The probability decreases linearly with each iteration, from an initial first probability of 0.08 to a final second probability of 0.02, reducing the likelihood of mutations disrupting the optimal solution in later iterations.

[0050] For the particle swarm optimization algorithm, the inertia weight corresponding to the first stage of the optimization process is determined as the first weight, the inertia weight of the second stage is determined as the second weight, and the inertia weight of the third stage is determined as the third weight. The first stage precedes the second stage, the second stage precedes the third stage, the first weight is greater than the second weight, and the second weight is greater than the third weight. For example, the particle size is set to 60, and each particle corresponds to one set of optimization variables; the initial velocity range is set to ±10% of the value range of the optimization variables (to reduce the possibility of the initial velocity being too large and jumping out of the feasible region). Then, the inertia weight is adaptively segmented. The first stage is the initial optimization stage (e.g., the first 30% of iterations): inertia weight ω=0.8-0.9 (strengthening global search and covering more feasible regions). The second stage is the middle optimization stage (30%-70% of iterations): ω=0.6-0.8 (balancing global exploration and local development). The third stage is the later optimization stage (the last 30% of iterations): ω=0.4-0.6 (focusing on local optima and accelerating convergence). The learning factors are set as individual learning factor c1=2.0 and global learning factor c2=2.0 (to adapt to the convergence characteristics of air conditioning system optimization).

[0051] For the simulated annealing algorithm, the cooling rate in the first stage of the optimization process is defined as the first rate, the cooling rate in the second stage as the second rate, and the cooling rate in the third stage as the third rate. The first rate is less than the second rate, and the second rate is less than the third rate. For example, the initial temperature and cooling strategy are as follows: Initial temperature T0 = 100 (calibrated based on the range of values ​​of the optimization variables), using exponential cooling with a dynamic cooling rate. The cooling rate in the first stage is α = 0.85 (rapidly reducing the temperature to cover a large feasible region). The cooling rate in the second stage is α = 0.9 (slowly reducing the temperature to refine the local search). The cooling rate in the third stage is α = 0.95 (stabilizing the temperature to avoid discarding the optimal solution). Then, a random perturbation method is used, with the perturbation amplitude of each optimization variable decreasing as the temperature decreases (e.g., initially ±8% of the value range, later decreasing to ±2%).

[0052] Based on the total energy consumption prediction value output by the target total energy consumption prediction model, an objective function for minimizing total energy consumption is constructed, and a penalty term is applied to the optimal solution that violates the two-layer constraint system. First, the objective function is defined, with the total energy consumption prediction value output by the target total energy consumption prediction model as the core, constructing a "total energy consumption minimization" objective function. For example, the objective function can satisfy Min F(X) = E_pred(X), where X=[x1, x2, x3, x4], x1 = chilled water outlet temperature, x2 = chilled water mass flow rate, x3 = cooling water inlet temperature, x4 = cooling water mass flow rate. Hard constraint penalty term (P_hard): If the optimization variable X violates the hard constraint, the penalty term is 20% of the current total energy consumption prediction value, formula: P_hard = 0.2×E_pred(X) (when X exceeds the hard constraint interval), otherwise P_hard=0. Soft constraint penalty term (P_soft): If the optimization variable X causes indoor temperature fluctuations > ±0.5℃, the penalty term is 10% of the current total energy consumption prediction value, calculated as: P_soft = 0.1 × E_pred(X) (when the fluctuation exceeds the limit), otherwise P_soft = 0. Final objective function: Min F'(X) = E_pred(X) + P_hard + P_soft.

[0053] Next, the target total energy consumption prediction model is embedded into the optimization process of multiple global optimization algorithms, using optimization variables as the optimization dimension. The optimization processes of multiple global optimization algorithms are started simultaneously, and a unified optimization termination condition is set. Specifically, the trained target total energy consumption prediction model is used as an "energy response calculator," embedded into the optimization processes of GA, PSO, and SA: for each set of optimization variable combinations X generated by the algorithm, the model output E_pred(X) is called and substituted into the objective function F'(X) to calculate the fitness value. Unified optimization parameters are configured: the maximum number of iterations is set to 100, and the initial feasible region is the hard constraint interval of the two-layer constraint system. A multi-threaded parallel computing approach is adopted, simultaneously starting the optimization processes of GA, PSO, and SA to reduce the inefficiency caused by serial computation. The optimal fitness value and average fitness value for each generation are recorded. If the fitness value does not improve for 10 consecutive generations, a "local optimum exit mechanism" is automatically triggered (e.g., GA increases the mutation probability by 50%, PSO resets 10% of particle positions, and SA increases the temperature by 10%). The optimization process terminates when any of the following conditions are met: the number of iterations reaches 100; or the total energy consumption change rate over 5 consecutive iterations is ≤0.5%. Upon termination, output the optimal fitness value and the corresponding combination of optimization variables for each algorithm type.

[0054] For each type of global optimization algorithm, optimization performance data corresponding to three dimensions—optimization accuracy, optimization speed, and optimization stability—are extracted. Specifically, the total energy consumption reduction rate (η_total) is calculated as follows: η_total = (E_base - E_opt) / E_base × 100%, where E_base is the baseline total energy consumption of the air conditioning system under unoptimized operating conditions (historical average for the same period), and E_opt is the optimal total energy consumption after algorithm optimization. An additional high-load operating condition energy consumption reduction rate (η_high) is extracted: operating condition data with air conditioning system cooling load ≥ 80% of rated cooling capacity are selected and calculated separately using the above formula, adapting to core energy-saving scenarios. The total time from startup to meeting the termination conditions for each type of algorithm is recorded, including model call time, fitness calculation time, and algorithm iteration time. The average iteration time is then calculated to reflect the algorithm's real-time response capability. The optimization operation is repeated 5 times for each type of algorithm (keeping the same constraints and initial parameters) to obtain 5 sets of total energy consumption reduction rates. Next, calculate the coefficient of variation (CV): CV = σ_η / μ_η × 100%, where σ_η is the standard deviation of the reduction rate of the 5 groups, and μ_η is the mean. The smaller the CV, the better the stability of the optimization.

[0055] Finally, the optimization performance data of each global optimization algorithm pair were correlated and integrated according to the dimensions of algorithm type, optimization accuracy, optimization speed, and optimization stability to obtain the optimization performance data corresponding to different global optimization algorithms. Specifically, a structured data matrix was constructed: using "algorithm type" as the row index and "optimization accuracy, optimization speed, and optimization stability" as the column index, and the extracted performance data was filled in. Key information in the optimization process (such as whether a local optimum escape mechanism was triggered and the type of termination condition) was labeled for each set of data to provide complete data support for subsequent multi-dimensional trade-off analysis.

[0056] The dual-layer constraint system avoids equipment overload and damage caused by optimization variables exceeding the equipment's rated range, while ensuring thermal comfort through indoor temperature fluctuation constraints, thus addressing the technical pain point of traditional optimization that prioritizes energy saving over safety and comfort. Differentiated optimization strategies customize parameters based on the core characteristics of three types of algorithms: GA enhances global exploration capabilities, PSO balances exploration and development, and SA avoids local optimum traps. Compared to single algorithms or general parameter settings, these strategies more comprehensively explore the optimal value space of optimization variables and improve convergence speed by 30%-50%. The objective function with a penalty term clearly defines the optimization direction, and the parallel optimization mechanism significantly shortens the total computation time, adapting to the real-time control requirements of dynamic air conditioning system operation and ensuring that the optimization scheme can quickly respond to changes in operating conditions. The extracted optimization performance data covers three core dimensions: accuracy, speed, and stability, and refines key sub-indicators such as energy-saving effect under high load conditions and average iteration time, providing a comprehensive and accurate data foundation for subsequent multi-dimensional trade-off analysis of model-algorithm objective combinations and reducing decision-making biases based on single indicators. The extraction and integration of optimization stability data quantifies the robustness of the algorithm, reduces the fluctuations in optimization effect caused by random fluctuations in the algorithm, and ensures that the air conditioning system can stably realize its energy-saving potential during long-term operation.

[0057] Step 104: Based on the optimized performance data, through a multi-dimensional trade-off analysis of prediction accuracy, optimization accuracy, and optimization speed, the objective weights are calculated using the entropy weight method, and a correction weight for optimization accuracy under high load conditions is assigned to determine the target combination of the target total energy consumption prediction model and the global optimization algorithm.

[0058] Multi-dimensional trade-off analysis refers to comprehensively considering four core dimensions: prediction accuracy (model capability), optimization accuracy (energy-saving effect), optimization speed (response efficiency), and optimization stability (operational reliability). Through quantitative calculation and hierarchical screening, it reduces decision-making bias caused by single-indicator guidance and achieves "multi-objective synergistic optimization." Prediction accuracy is the core performance indicator of the target total energy consumption prediction model. It quantifies the degree of error between the model's predicted value and the actual energy consumption, serving as the "basic input quality assurance" for the optimization algorithm. Optimization accuracy is the core performance indicator of the global optimization algorithm. It quantifies the reduction rate of total energy consumption relative to the baseline operating condition after optimization (including special indicators for high-load operating conditions), directly reflecting the energy-saving effect. Optimization speed is the response efficiency indicator of the global optimization algorithm. It quantifies the total time taken from algorithm startup to meeting the termination conditions, adapting to the real-time control needs of the dynamic operation of the air conditioning system. The target combination is the optimal combination of "prediction model-optimization algorithm" determined after multi-dimensional trade-offs and two-layer screening, which can simultaneously meet the global optimization requirements of "accurate prediction, strong energy saving, fast speed, and stable operation." By conducting multi-dimensional assessments of prediction, optimization, speed, and stability, we break away from the traditional single-minded decision-making model that only considers energy-saving effects. This ensures that the target combination takes into account accuracy, effectiveness, real-time performance, and reliability, and adapts to the complex operational needs of air conditioning systems.

[0059] Specifically, candidate combinations of the target total energy consumption prediction model and the global optimization algorithm are first obtained, and the prediction accuracy, optimization accuracy, and optimization speed are associated with each candidate combination to obtain the mapping relationship between the candidate combination and multi-dimensional indicators. A candidate combination refers to all feasible pairings of "target total energy consumption prediction model + global optimization algorithm," covering combination types with different modeling characteristics (linear / nonlinear) and optimization characteristics (global exploration / fast convergence / avoidance of local optima), serving as the basic sample pool for decision-making. For example, based on the above four types of target total energy consumption prediction models and three types of global optimization algorithms, 12 non-repeating feasible candidate combinations can be generated, ensuring coverage of all pairings of linear / nonlinear models with algorithms of different optimization characteristics, and eliminating invalid combinations with conflicting model and algorithm characteristics (such as redundant pairings of linear models + strongly nonlinear algorithms).

[0060] The index mapping relationship refers to the one-to-one correspondence between candidate combinations and corresponding performance indicators. It is the basic data structure for subsequent quantitative analysis, ensuring the traceability and accuracy of the "combination-indicator" relationship. For example, prediction accuracy: extract the MAPE value of each target model from the model validation results in step 102 (e.g., MAPE = 5.2% for MLR, MAPE = 3.8% for ANN); optimization accuracy: extract the "total energy consumption reduction rate" and "high load condition energy consumption reduction rate" for each algorithm from the performance data in step 103 (e.g., total reduction rate of GA = 12%, high load reduction rate = 15%); optimization speed: extract the total optimization time for each algorithm from the performance data in step 103 (e.g., optimization time of PSO = 8s, optimization time of SA = 12s); construct a mapping table: form a structured mapping table with "candidate combination name" as the row and "prediction accuracy (MAPE), total energy consumption reduction rate, high load reduction rate, optimization time, and optimization stability coefficient of variation" as the columns, ensuring that the index data of each combination is unique and traceable.

[0061] It comprehensively covers all feasible combinations and core indicators, minimizing the omission of high-quality combinations or key performance dimensions, and providing complete data support for subsequent analysis. The structured mapping table makes the correspondence between "combination-indicator" clear and traceable, reducing data confusion and minimizing subsequent calculation errors.

[0062] After normalizing the indicators for each dimension, the entropy weight method can be used to calculate the objective weight of each indicator. Normalization converts performance indicators with different dimensions and value ranges into standardized data within the [0,1] interval, eliminating the impact of dimensional differences on weight calculation and score aggregation, ensuring comparability between indicators. The entropy weight method is an objective weight allocation method based on the dispersion of indicator data. Its core logic is that "the higher the indicator's discriminative power (the lower the information entropy), the greater its contribution to decision-making, and the higher its weight," avoiding weight bias caused by subjective experience. For example, first, prediction accuracy, optimization speed, and optimization accuracy are positively transformed, and the Min-Max normalization method is used to map all positive indicators to the [0,1] interval, followed by outlier handling. Then, the information entropy of each normalized indicator is calculated; information entropy is an objective weight allocation method that quantifies the dispersion of indicator data. The indicator difference coefficient is calculated; the higher the indicator's discriminative power, the greater its weight. After calculating the basic weights, the weight system is corrected. For example, the optimization accuracy under high load conditions is given an additional corrected weight of 0.2, and the basic weights are scaled proportionally to form a complete weight system.

[0063] Then, based on the objective weights of the indicators in each dimension, a weighted summation method is used to calculate the comprehensive score of each candidate combination. For example, the comprehensive score = Σ(indicator normalized value × corresponding weight), specifically decomposed as: Comprehensive score = (prediction accuracy normalized value × prediction accuracy weight) + (optimization accuracy normalized value × optimization accuracy weight) + (optimization speed normalized value × optimization speed weight) + (high load optimization accuracy normalized value × high load correction weight). The candidate combination data is then substituted into the calculation. Taking a candidate combination "ANN+GA" as an example: Normalized prediction accuracy = 0.92, weight = 0.3 → contribution value = 0.92 × 0.3 = 0.276; Normalized optimization accuracy = 0.85, weight = 0.4 → contribution value = 0.85 × 0.4 = 0.34; Normalized optimization speed = 0.7, weight = 0.1 → contribution value = 0.7 × 0.1 = 0.07; Normalized high-load optimization accuracy = 0.9, weight = 0.2 → contribution value = 0.9 × 0.2 = 0.18; Overall score = 0.276 + 0.34 + 0.07 + 0.18 = 0.866. Calculate the overall score of all 12 candidate combinations using the above formula, sort them from highest to lowest score, and form a ranking table of combination performance. Transforming multi-dimensional, non-comparable indicators into a single comprehensive score makes the merits and demerits of different combinations readily apparent, reducing decision-making complexity. Through weighted scoring, the importance of indicators such as prediction accuracy, optimization accuracy, and optimization speed is reflected in the score, ensuring that the comprehensive score reflects the need for "multi-objective synergistic optimization." Score ranking provides a clear priority for subsequent screening, facilitating the rapid identification of high-performing combinations.

[0064] Finally, candidate combinations with a comprehensive score greater than or equal to the set score are selected, and the combination with the highest comprehensive score under high-load conditions and a stability coefficient of variation less than or equal to the set coefficient of variation is chosen as the target combination. The selection threshold can include the set score and the set coefficient of variation. The set score is a comprehensive score threshold based on actual engineering needs (such as energy-saving targets and response speed requirements), used to quickly eliminate low-quality combinations that do not meet performance standards. The set coefficient of variation is an optimization stability threshold based on the stability requirements of the air conditioning system, used to select combinations with small fluctuations in results and reliable long-term operation. The comprehensive score under high-load conditions is a comprehensive score specifically calculated for core high-load scenarios, strengthening the performance weight during core energy-saving periods to ensure that the target combination performs optimally under critical scenarios.

[0065] For example, from 12 combinations, combinations with a comprehensive score greater than or equal to 0.8 are selected, assuming four combinations are retained: "ANN+GA", "RF+GA", "ANN+PSO", and "RF+PSO". The first step in the selection process is to verify the coefficient of variation of the optimization stability of the initially selected combinations. For example, combinations with a coefficient of variation > 0.05 are eliminated (assuming "RF+PSO" is eliminated, leaving 3 combinations). The second step in the selection process is to calculate the "comprehensive score under high load conditions" of the remaining combinations (considering only the weighted score of prediction accuracy, high load optimization accuracy, and optimization speed), and select the combination with the highest score (assuming "ANN+GA" has a high load score of 0.89, which is the highest). Finally, "ANN+GA" (artificial neural network model + genetic algorithm) is determined as the target combination of the target total energy consumption prediction model and the global optimization algorithm.

[0066] The dual-layer screening mechanism ensures both the overall performance of the combination and operational stability, while also reinforcing the advantages of core scenarios (highest scores under high load) and reducing combination bias caused by single-dimensional screening. The optimized stability threshold reduces the selection of combinations that perform well in a single instance but exhibit large long-term fluctuations, ensuring that the air conditioning system consistently delivers its energy-saving potential over long-term operation. The screening threshold is linked to engineering requirements, with specialized screening for high-load conditions tailored to the core operating scenarios of commercial building air conditioning. Through a layered logic of initial and refined screening, the candidate pool is gradually narrowed down, ultimately focusing on a single target combination, improving decision-making efficiency and accuracy.

[0067] Step 105: Combine the predicted values ​​of interference factors, embed the target combination into the global optimization architecture of the air conditioning system, and output the target set value of the air conditioning system after predictive modeling and global optimization corresponding to the target combination.

[0068] The target setpoint is the optimal setpoint. The predicted values ​​of disturbance factors are the prediction results of dynamic external variables affecting the operation of the air conditioning system. The core factors are the cooling load (indoor cooling demand) and the outdoor wet-bulb temperature (outdoor meteorological conditions), which are key inputs for dynamic optimization. The selected target combination (i.e., the prediction model and optimization algorithm) is embedded into the actual global optimization architecture. Combined with accurate predictions of cooling load and outdoor wet-bulb temperature, and through a closed-loop process of real-time data acquisition, dynamic prediction, boundary updates, rolling optimization, and setpoint optimization, a safe, adaptable, and energy-saving target setpoint for the air conditioning system is output, achieving globally optimal control under dynamic operating conditions.

[0069] Specifically, the predicted values ​​of disturbance factors for the global optimization architecture adapted to the air conditioning system are obtained. These predicted values ​​can include cooling load predictions and outdoor wet-bulb humidity predictions. The disturbance factor prediction employs a combination of Long Short-Term Memory (LSTM) networks and temporal smoothing. Inputting historical time-series data of cooling load and outdoor wet-bulb temperature for the previous 30 minutes, the system predicts time-period values ​​(e.g., one data point every 10 minutes) within a set optimization period (10 minutes to 1 hour). The prediction is only activated when the root mean square error of the predicted value is less than or equal to 5; otherwise, it switches to the historical average value for the same period to ensure input reliability. The combined prediction method balances temporal trend capture and fluctuation suppression, reducing the prediction bias of a single model. Predicting future changes in disturbance factors in advance makes the optimization strategy forward-looking, reducing passive adjustments and adapting to dynamic operating conditions.

[0070] Real-time data collection of the air conditioning system's current operating conditions is crucial. This data reflects the system's current status and provides a baseline for prediction. The system collects current operating data (measured chilled / cooling water temperature, flow rate, etc.) at intervals consistent with the optimization step size. The current operating data and predicted values ​​of interference factors are input into the target total energy consumption prediction model to predict the time-period energy consumption for a future optimization period. This prediction serves as the objective function of the global optimization algorithm. The time-period energy consumption prediction results represent the energy consumption forecast for each time point within the future period, allowing optimization to be tailored to time-period differences. Predicting the time-period energy consumption within the future optimization period directly serves as the input to the global optimization algorithm's objective function (minimizing total energy consumption). Based on the prediction results using current and future data, the optimization algorithm's objective function better aligns with actual operating conditions, reducing optimization failures caused by static objectives. Establishing a direct link between prediction output and optimization input ensures the smooth operation of the system with the combined objectives.

[0071] Then, based on the dynamic changes in the predicted values ​​of disturbance factors, the constraint boundaries of the optimization variables are updated in real time. For example, the constraints of the optimization variables are adjusted according to the changes in the predicted values ​​of disturbance factors, with the update frequency synchronized with the optimization step size, ensuring that the constraint boundaries dynamically match the operating conditions. This updates ensure that the minimum rated parameters of the equipment are met, reducing the possibility of constraint failures, guaranteeing the feasibility of optimization, and allowing the dynamically constrained optimization variables to better reflect real-time operating conditions, thus unlocking greater energy-saving potential.

[0072] Next, with the goal of minimizing energy consumption over time periods, and in conjunction with the updated constraint boundaries, the optimization variables are optimized. During the optimization process, the energy consumption prediction values ​​fed back in real time from the target total energy consumption prediction model are invoked, forming a closed-loop rolling optimization process of optimization, feedback, and adjustment. Rolling optimization is an optimization method that repeatedly executes the optimization process with a fixed optimization step size, dynamically adapting to changes in operating conditions. The prediction model is continuously invoked during the optimization process to ensure that energy consumption assessment and variable adjustment are synchronized. The closed-loop feedback mechanism enables the algorithm to correct the optimization direction in real time, reducing optimization failures caused by prediction bias. The rolling mode addresses the long time delay and nonlinearity problems of air conditioning systems, achieving continuous optimization under dynamic operating conditions.

[0073] The optimization process terminates when the number of iterations per single step is greater than or equal to a set number, or when the energy consumption change rate after a set number of consecutive iterations is less than or equal to a set change rate. The number of iterations per single step refers to the number of iterations the algorithm performs within each optimization step. The energy consumption change rate is the percentage difference in energy consumption prediction between two adjacent iterations, reflecting the degree of convergence. Both the set number of iterations and the set number of iterations are thresholds for determining the optimization termination condition. For example, the set number of iterations can be 30, and the set number of iterations can be 3. Triggering either condition will stop the optimization at the current step.

[0074] The optimization results for each optimization step are extracted, and the time-series setpoints for each time period within a future optimization period are integrated according to the time series. These time-series setpoints contain the specific values ​​of the optimized variables. The time-series setpoints are the set of optimal values ​​of the optimized variables arranged in a time series, representing the control command sequence of the air conditioning system. By concatenating the time-series setpoints for each time period within a future optimization period in chronological order, the specific values ​​of all optimized variables can be included. Integrating discrete optimization results into a continuous time-series sequence can adapt to the continuous operation requirements of the air conditioning system. Structured time-series data provides a foundation for engineering adaptation.

[0075] Finally, linear change processing is performed on the timing setpoints, and outliers exceeding the rated operating range of the air conditioning equipment are removed and replaced with safe and feasible values. The target setpoints of the air conditioning system are then output. Linear change processing smooths the differences between setpoints in adjacent time periods, reducing sudden changes in equipment parameters. For example, the variation range of adjacent step setpoints can be constrained (temperature ≤ 0.5℃, flow rate ≤ 5%). Safe and feasible values ​​refer to variable values ​​that conform to the rated operating range of the equipment and ensure system stability. Outliers exceeding the rated operating range are replaced with the most recent safe and feasible values, and the output target setpoints can be directly applied to the air conditioning equipment. This reduces equipment shocks and overloads caused by sudden parameter changes and extends equipment lifespan. The setpoints after smoothing and error correction can be directly executed, solving the problem of the disconnect between theoretical optimization and engineering implementation.

[0076] Please refer to Figure 2, which illustrates a schematic diagram of a global optimization architecture for an air conditioning system provided in this application embodiment. In the figure, squares represent elements of the architecture, solid lines point to constituent elements, and dashed lines point to the application objects of the elements (methods). The architecture consists of three main parts: a total energy consumption prediction model, rolling optimization, and modeling and optimization methods: 1) Determining key factors affecting the system's total energy consumption through sensitivity analysis, including disturbance variables and optimization variables. Disturbance variables consist of cooling load and outdoor meteorological parameters, while optimization variables are optimizable operating parameters of the system, such as chilled water outlet temperature and cooling water inlet temperature; 2) Establishing a prediction model for the system's total energy consumption and its disturbance variables based on machine learning modeling methods; 3) The rolling optimization method uses global optimization algorithms, such as GA and PSO algorithms; 4) Optimizing the optimization variables in the total energy consumption prediction model based on the rolling optimization method to minimize the system's total energy consumption; 5) Instantiation study of the global optimization architecture in a real air conditioning system.

[0077] In general air conditioning systems, local equipment often employs stable feedback control methods. For example, the fan speed on the terminal heat exchange side is controlled based on PID to meet cooling requirements. In contrast, some important parameters in the system need to be set based on experience, such as the chilled water outlet temperature and its flow rate (Tchws and mchw), and the cooling water inlet temperature and its flow rate (Tcws and mcw). Although the qualitative method based on experience can better guarantee the cooling capacity supply at the terminal, it is difficult to obtain the optimal energy efficiency of the system. Accordingly, this application embodiment takes the minimum total energy consumption of the system as the optimization objective and optimizes the setting of these parameters as shown in the following formula (1).

[0078] (1)

[0079] In the formula, TECPM (Total Energy Consumption Prediction Model) is the system's total energy consumption prediction model; Tchws,min and Tchws,max are the minimum and maximum values ​​of the chilled water outlet temperature, respectively, in °C; mchw,min and mchw,max are the minimum and maximum values ​​of the chilled water mass flow rate, respectively, in kg / s; Twet(t) represents the outdoor wet-bulb temperature at time t, in °C; ΔTapp,min and ΔTapp,max represent the minimum and maximum approximations (°C) of the cooling tower outlet temperature, respectively, generally taken as 1°C-5°C; mcw,min and mcw,max are the minimum and maximum values ​​of the cooling water mass flow rate, respectively, in kg / s; CL is the cooling load, in W; and Twet is the outdoor wet-bulb temperature, in °C.

[0080] As shown in Figure 3, Figure 3 is a schematic diagram of a rolling optimization process provided in an embodiment of this application. It reflects the time sequence flow of rolling time-domain optimization for global optimization of the air conditioning system and demonstrates the time-by-time closed-loop control logic of dynamic disturbance input-prediction+optimization-optimization variable output, corresponding to the core process of combining disturbance factor prediction and rolling optimization to output target setpoints in the previous technical solution. Based on the energy consumption optimization problem given by equation (1), the rolling optimization process of this embodiment of the application is constructed to ensure continuous optimization. The input disturbance variables are dynamic external factors affecting the operation of the air conditioning system, specifically including (t): The predicted cooling load at time t, reflecting the indoor cooling demand; : The predicted outdoor wet-bulb temperature at any given time reflects the impact of outdoor weather conditions on air conditioning heat dissipation. First, an appropriate optimization step size is set. This allows for better capture of the system's dynamic characteristics. However, a smaller optimization step size is not always better; an excessively small step size requires significant computational memory and can easily lead to model instability. The optimization step size should cover the duration of fluctuations caused by system inertia, typically ranging from 10 minutes to 1 hour. The step size is dynamically updated with the time step, covering subsequent time periods (…). + , +2 ... + The predicted values ​​of disturbance factors correspond to each time period. Furthermore, in the Total Energy Consumption Prediction Model (TECPM), some variables are disturbance variables that need to be predicted in advance, while others are optimization variables (corresponding to the controlled variables of the system), which are determined through optimization. Therefore, at each optimization step ( In each iteration of the TECPM, the values ​​of the disturbance variables (CL, Twet) are updated to ensure the model adapts to the changing environment in real time. Then, the optimizer optimizes the optimized variables in the TECPM to obtain the optimal system settings. Optimized variables may include the chilled water outlet temperature setpoint. chilled water mass flow rate setpoint Cooling water inlet temperature setpoint Cooling water mass flow rate setpoint The optimized operating parameters for each time period are dynamically updated with a fixed time step. Closed-loop control is performed periodically, using a fixed time step as the unit. For example, input... Interference variables during the time period [ (t), The energy consumption for this period is predicted by TECPM(t), and then optimized by the Optimizer to output the result. Optimization variables for the time period. Input + Disturbance variables at time [ ( + ), ], call TECPM( + ) and Optimizer, output + Optimization variables for different time periods. The process of repeated input-prediction-optimization-output covers... arrive + Throughout the entire optimization cycle, it achieves real-time adaptation to dynamically changing cooling loads and outdoor weather conditions. The entire process moves forward in a rolling fashion, solving the static optimization failure problem caused by the strong coupling and time-varying nonlinear characteristics of the air conditioning system, and ensuring optimal energy consumption and stable operation of the system at all times.

[0081] Figure 4 is a schematic diagram of a non-convex surface of an objective function in an embodiment of this application. As shown in Figure 4, it consists of optimization variables and time-varying disturbance variables. The colored three-dimensional surface in the figure represents the value distribution of the objective function (total energy consumption). The two bottom dimensions of the surface correspond to the optimization variables and disturbance variables. The height dimension of the surface corresponds to the value of the objective function (i.e., the magnitude of total energy consumption), with the color from blue to red indicating a value from low to high (blue for low energy consumption and red for high energy consumption). The red dashed circle outlines two key nodes: the local minimum and the global minimum. The local minimum is the "relative lowest point" in a certain local region of the surface (such as the valley in the blue-green region), but it is not the lowest value of the entire surface. If a local optimization algorithm (such as gradient descent) is used, it is easy to get stuck in this region, resulting in the optimization result not being the true energy-optimal. The global minimum is the "absolute lowest point" of the entire surface (such as the deep valley in the dark blue region), corresponding to the "optimal solution of total energy consumption" of the air conditioning system under the current operating conditions, which is also the target that the "global optimization algorithm (genetic algorithm, particle swarm optimization, etc.)" in the previous technical solution seeks to find. When the disturbance variable changes, the optimization surface also changes, causing the minimum point to shift. This shows that the optimal value of the optimization variable differs under different disturbance variables, highlighting the importance of predicting the disturbance variable. Furthermore, when the optimization range is narrow, the optimal value obtained may be a local minimum; however, when the optimization range is expanded, the global minimum may be found. This demonstrates the crucial importance of accurately defining the optimization range for optimizing system energy consumption.

[0082] Figure 5 is a schematic diagram of a global optimization architecture for an air conditioning system provided in a specific embodiment of this application. As shown in Figure 5, the prediction model and optimization algorithm, as the core of the global optimization architecture, clearly present the entire technical logic from data preprocessing, prediction model construction and optimization, algorithm optimization, model and algorithm matching, and final setpoint output. Their performance is related to the optimization capability of the architecture. In order to further enhance the optimization performance of the architecture, this embodiment of the application proposes a performance improvement method for the global optimization architecture from two aspects: prediction model performance improvement method, Part 1 (prediction model and optimization algorithm matching method) and Part 2 (setpoint optimization implementation link). Part 1 is the core component construction and matching link of global optimization, which is responsible for building a high-precision prediction model, screening efficient optimization algorithms, and finding the best combination of the two. Specifically, it may include 1) after determining the key variables affecting the total energy consumption of the air conditioning system through sensitivity analysis, preprocessing the resulting operating data to obtain the training data of the model. Specifically, seven types of core operating / environmental parameters are input as basic data, including the chilled water outlet temperature setpoint. chilled water mass flow rate setpoint Cooling water inlet temperature setpoint Cooling water mass flow rate setpoint The data includes: cooling load (CL), outdoor wet-bulb temperature (Twet), and total system power (P, representing total energy consumption). Preprocessing operations such as outlier removal and normalization are performed on the collected historical data to ensure data integrity and provide reliable input for subsequent model training. 2) A total energy consumption prediction model is constructed based on different machine learning methods, and grid search and cross-validation methods are used to improve the performance of the prediction model. The total energy consumption model can include MLR, ANN, SVR, and RF models. 3) Different global optimization algorithms are applied to optimize the variables in the total energy consumption model to obtain the performance of different optimization algorithms. The global optimization algorithms include GA, PSO, and SA algorithms. 4) Based on the trade-off analysis of prediction accuracy, optimization accuracy, and optimization speed, the optimal combination of the prediction model and optimization algorithm is obtained. Part 2 is the implementation stage of global optimization. The optimal combination obtained in Part 1 can be used to output the setpoints that can be directly controlled by the air conditioning system. This can include: 1) Inputting the core disturbance factors for the predicted future period: cooling load (CLt+1) and outdoor wet-bulb temperature (Twet,t+1), providing forward-looking input for dynamic optimization; 2) Using the optimal prediction model and optimal optimization algorithm selected in Part 1 as the core processing unit, inputting the predicted values ​​of the disturbance factors, and performing prediction and optimization calculations; 3) Outputting the core operating parameter setpoints of the air conditioning system, including: chilled water outlet temperature (Tchws,t+1), cooling water inlet temperature (Tcws,t+1), cooling water mass flow rate (mcws,t+1), and chilled water mass flow rate (mchw,t+1). The optimal combination is applied to the global optimization architecture, and by inputting the predicted values ​​of the disturbance factors, the optimal system setpoints are output after optimization. Starting with basic data processing, we first construct and optimize the prediction model and deploy the global optimization algorithm. Then, through multi-dimensional analysis, we find the best combination of "model-algorithm". Finally, we combine dynamic disturbance prediction to output executable system setpoints, realizing a closed loop of "accurate prediction → efficient optimization → dynamic control", which solves the problem of global energy consumption optimization under the "multi-variable coupling and dynamic changes in operating conditions" of air conditioning systems.

[0083] Figure 6 is a schematic diagram of the application of a global optimization method for an air conditioning system in an embodiment of this application. As shown in Figure 6, this schematic diagram is a smart optimization control architecture diagram of an air conditioning system based on digital twin technology. (a) is a conceptual diagram of optimization control, and (b) is a diagram of the actual physical system, which fully presents the closed-loop control logic of virtual twin model - actual physical system - intelligent monitoring.

[0084] (a) illustrates an iterative optimization framework of prediction-execution-feedback, including a twin prediction system (corresponding to a load-energy consumption dual prediction model and a cross-domain coupling mechanism) and an actual physical model (corresponding to an actual operating air conditioning system). Operational data of the actual physical system (such as equipment parameters and environmental parameters) is acquired through "data acquisition" and input into the twin prediction model. Based on this data, the model executes algorithms such as cross-domain coupling, load-energy consumption prediction, and multi-objective regulation to generate optimized control strategies. After the strategies are fed back to the actual physical system for execution, new operational data is again fed back to the twin prediction model, completing "feedback optimization," corresponding to the iterative optimization process in step 105 above, achieving continuous improvement in control accuracy.

[0085] (b) illustrates the hardware components, data acquisition links, and control flow of the actual air conditioning system, corresponding to the actual implementation path of the control methods described earlier. The hardware components include chillers (providing the cooling source), chilled water pumps (delivering chilled water to the terminals), cooling water pumps (delivering cooling water to the chillers), cooling towers (cooling the high-temperature cooling water outlet from the chillers), and indoor air conditioning terminals (fan coil units, responsible for heat and mass exchange within the rooms). The data acquisition link involves collecting key operational data, such as chilled / cooled water flow rates and chiller / pipeline water temperatures, via flow sensors (labeled "P") and temperature sensors (labeled "T") on the equipment. Simultaneously, "weather" data (outdoor temperature and humidity, wet-bulb temperature, etc., corresponding to the interference variables mentioned earlier) is also collected. This data is uniformly transmitted to the monitoring controller (the module labeled "brain" in the upper right corner of the diagram). The monitoring execution chain includes: the monitoring controller calls the "twin prediction model," which, based on the collected operational and weather data, executes the algorithms described earlier, such as "cross-domain coupling mechanism construction → load-energy consumption dual prediction → multi-objective control sequence generation," and outputs optimized control signals ("dashed lines" in the diagram); the control signals are directly sent to each device: adjusting the operating frequency of cooling water pumps / chilled water pumps, the start / stop status of cooling towers, and the operating parameters of chillers, ultimately controlling the cooling / heating effect of fan coil units in the room to achieve the multi-objective requirements of "load matching, energy consumption optimization, and stable operation." Closed-loop feedback includes: new operational data after the devices execute control signals is transmitted back to the monitoring controller via flow / temperature sensors, entering the next cycle of "data acquisition → model calculation → control execution," corresponding to the iterative process described earlier, continuously adapting to the actual operating status.

[0086] In this way, the methods such as "cross-domain coupling, predictive calibration, and multi-objective control" abstracted in the embodiments of this application can be mapped to the hardware link of "sensor acquisition → controller calculation → device execution → data feedback" in the actual system, clearly presenting the implementation logic of optimized control.

[0087] Based on the same inventive concept, embodiments of this application also provide a computer-readable storage medium storing a program that can be loaded by a processor and executed as any of the global optimization methods for an air conditioning system in embodiments of this application.

[0088] Those skilled in the art will understand that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be achieved. In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the program can also be stored in a server, another computer, disk, optical disk, flash drive, or external hard drive, etc., and can be downloaded or copied to the memory of a local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be achieved.

[0089] The above examples illustrate this application only to aid understanding and are not intended to limit its scope. Those skilled in the art to which this application pertains can make various simple deductions, modifications, or substitutions based on the ideas presented.

Claims

1. A global optimization method for an air conditioning system, characterized in that, The method includes: collecting historical operating data of the air conditioning system and corresponding indoor and outdoor environmental data; performing global sensitivity analysis using the Morris screening method to determine the target variable affecting the total energy consumption of the air conditioning system; obtaining training data based on the target variable; the global sensitivity analysis is used to quantify the main effects of a single variable and the coupling effects between variables; inputting the training data into a pre-constructed training architecture of multiple initial total energy consumption prediction models for the air conditioning system; using a differentiated grid search method to determine the optimal hyperparameters of each initial total energy consumption prediction model; verifying the generalization ability of the initial total energy consumption prediction model; obtaining an optimized target total energy consumption prediction model; based on the preset value range of the optimization variable and associated with the indoor thermal comfort requirements of the air conditioning system, constructing a two-layer constraint system with a penalty term; and using a global optimization algorithm to optimize the optimization variable in the target total energy consumption prediction model to obtain different... The optimization performance data corresponding to the global optimization algorithm is obtained; the candidate combinations of the target total energy consumption prediction model and the global optimization algorithm are obtained, and the prediction accuracy, optimization accuracy and optimization speed are associated with each candidate combination to obtain the mapping relationship between the candidate combination and multi-dimensional indicators; after normalizing the indicators of each dimension, the objective weight of the indicator of each dimension is calculated using the entropy weight method; based on the objective weight of the indicator of each dimension, the comprehensive score of each candidate combination is calculated using a weighted summation method; the candidate combinations with a comprehensive score greater than or equal to a set score are screened, and the combination with the highest comprehensive score under high load conditions and an optimization stability variation coefficient less than or equal to a set variation coefficient is selected as the target combination; combined with the predicted value of interference factors, the target combination is embedded into the global optimization architecture of the air conditioning system, and after prediction modeling and global optimization corresponding to the target combination, the target set value of the air conditioning system is output.

2. The global optimization method for an air conditioning system according to claim 1, characterized in that, The process involves collecting historical operating data of the air conditioning system and corresponding indoor and outdoor environmental data, performing global sensitivity analysis using the Morris screening method to determine the target variable affecting the total energy consumption of the air conditioning system, and obtaining training data based on the target variable. This includes: preprocessing the historical operating data and the indoor and outdoor environmental data; selecting a set of candidate variables strongly correlated with the operation of the air conditioning system, using the total system power as the response index for total energy consumption; setting the number of variable levels adapted to the multivariate coupling characteristics of the air conditioning system and the estimated number covering typical operating conditions of the air conditioning system, and constructing the sampling trajectory of the variables; calculating the global sensitivity index of each candidate variable in the candidate variable set, whereby the global sensitivity index is used to quantify the influence of a single variable and the coupling effect between variables on total energy consumption; dynamically setting a sensitivity index threshold based on the fluctuation confidence interval of the historical total energy consumption of the air conditioning system, and selecting candidate variables with a global sensitivity index greater than or equal to the sensitivity index threshold as the target variable; and dividing the target variable into a training set and a validation set according to the time interval ratio based on the temporal continuity of the historical operating data.

3. The global optimization method for an air conditioning system according to claim 2, characterized in that, The target variables include chilled water outlet temperature, chilled water mass flow rate, cooling water inlet temperature, cooling water mass flow rate, cooling load, outdoor wet-bulb temperature, and total power of the storage medium. The preprocessing of the historical operating data and the indoor and outdoor environmental data includes: for the chilled water outlet temperature, cooling water inlet temperature, and outdoor wet-bulb temperature, using... The criteria are used to identify and remove outliers; for the chilled water mass flow rate and the cooling water mass flow rate, the original cumulative flow rate data is converted into unit time flow rate data based on the sampling time interval and the original cumulative flow rate data; for the cooling load and the total system power, dimensional normalization processing is performed to map the values ​​to the [0,1] interval.

4. The global optimization method for an air conditioning system according to claim 2, characterized in that, The process involves inputting the training data into a pre-constructed training architecture for multiple initial total energy consumption prediction models of the air conditioning system. A differentiated grid search method is used to determine the optimal hyperparameters for each initial total energy consumption prediction model, and the generalization ability of the initial total energy consumption prediction models is verified to obtain an optimized target total energy consumption prediction model. This includes: pre-constructing initial sub-models corresponding to a multiple linear regression model, an artificial neural network model, a support vector regression model, and a random forest model to obtain initial total energy consumption prediction models; designing a differentiated grid search hyperparameter space for each initial sub-model; using the mean square error of total energy consumption prediction in the training set as the objective function, simultaneously performing a grid search on multiple initial sub-models, traversing the hyperparameter space of each initial sub-model, and selecting the hyperparameter combination with the smallest mean square error of total energy consumption prediction as the target hyperparameter of the initial sub-model. The validation set is divided into multiple sub-validation sets according to the operating conditions of the air conditioning system. Hierarchical cross-validation is performed on each initial sub-model equipped with the target hyperparameters. The mean square error variation coefficient of total energy consumption prediction for each initial sub-model in each sub-validation set is calculated. The operating conditions include high load, medium load, and low load conditions. If there exists an initial sub-model whose mean square error variation coefficient of total energy consumption prediction in all sub-validation sets is less than or equal to a set variation coefficient, then the generalization ability of the initial sub-model is deemed qualified. Among the initial sub-models with qualified generalization ability, the initial sub-model with the smallest mean square error of total energy consumption prediction in the training set and the smallest prediction error in the sub-validation set under high load conditions is selected as the target sub-model. Based on the target sub-model, the initial sub-model is optimized to obtain the target total energy consumption prediction model.

5. The global optimization method for an air conditioning system according to claim 4, characterized in that, The step of designing a differentiated grid search hyperparameter space for each of the initial sub-models includes: for the multiple linear regression model, setting the hyperparameter space as the regularization coefficient range; for the artificial neural network model, setting the hyperparameter space as the range of hidden layer node numbers and the learning rate range; for the support vector regression model, setting the hyperparameter space as the kernel function type and the kernel function parameter range; and for the random forest model, setting the hyperparameter space as the range of decision tree number and the maximum depth range.

6. The global optimization method for an air conditioning system according to claim 1, characterized in that, Based on the preset range of values ​​for optimization variables and in association with the indoor thermal comfort requirements of the air conditioning system, a two-layer constraint system with a penalty term is constructed. A global optimization algorithm is used to optimize the optimization variables in the target total energy consumption prediction model, obtaining optimization performance data corresponding to different global optimization algorithms. This includes: setting a safe range of values ​​for the optimization variables based on the rated operating parameters of the air conditioning equipment, and in association with the indoor thermal comfort requirements of the air conditioning system, adding constraints on indoor temperature fluctuations to obtain a two-layer constraint system; designing differentiated optimization strategies corresponding to multiple types of global optimization algorithms, including genetic algorithms, particle swarm optimization algorithms, and simulated annealing algorithms; based on the target total energy consumption... The total energy consumption prediction value output by the energy consumption prediction model is used to construct an objective function that minimizes the total energy consumption, and a penalty term is applied to the optimization solution that violates the two-layer constraint system. The objective total energy consumption prediction model is embedded into the optimization process of multiple global optimization algorithms, with the optimization variable as the optimization dimension. The optimization process of multiple global optimization algorithms is started simultaneously, and a unified optimization termination condition is set. For each type of global optimization algorithm, the optimization performance data corresponding to the three dimensions of optimization accuracy, optimization speed, and optimization stability are extracted. According to the dimensions of algorithm type, optimization accuracy, optimization speed, and optimization stability, the optimization performance data of each type of global optimization algorithm are associated and integrated to obtain the optimization performance data corresponding to different global optimization algorithms.

7. The global optimization method for an air conditioning system according to claim 6, characterized in that, The differentiated optimization strategies designed for the various global optimization algorithms include: for the genetic algorithm, retaining a predetermined percentage of the best individuals from the previous generation, dynamically adjusting the crossover probability based on the population fitness variance, and decreasing the mutation probability from a first probability to a second probability with each iteration, wherein the first probability is greater than the second probability; for the particle swarm optimization algorithm, determining the inertia weight corresponding to the first stage of the optimization process as the first weight, the inertia weight corresponding to the second stage as the second weight, and the inertia weight corresponding to the third stage as the third weight, wherein the first stage precedes the second stage, the second stage precedes the third stage, the first weight is greater than the second weight, and the second weight is greater than the third weight; for the simulated annealing algorithm, determining the cooling rate corresponding to the first stage of the optimization process as the first rate, the cooling rate corresponding to the second stage as the second rate, and the cooling rate corresponding to the third stage as the third rate, wherein the first rate is less than the second rate, and the second rate is less than the third rate.

8. The global optimization method for an air conditioning system according to claim 1, characterized in that, Combining the predicted values ​​of interference factors, the target combination is embedded into the global optimization architecture of the air conditioning system. After prediction modeling and global optimization corresponding to the target combination, the target setpoint of the air conditioning system is output, including: obtaining the predicted values ​​of interference factors adapted to the global optimization architecture of the air conditioning system, the predicted values ​​of interference factors including the predicted values ​​of cooling load and outdoor wet-bulb humidity, the prediction of interference factors adopts a combination of LSTM and time-series smoothing; real-time collection of the current operating data of the air conditioning system, inputting the current operating data and the predicted values ​​of interference factors into the target total energy consumption prediction model, predicting the time-period energy consumption prediction value of the future set optimization period, as the objective function of the global optimization algorithm; based on the dynamic changes of the predicted values ​​of interference factors, the target setpoint is then calculated. The constraint boundaries of the optimization variables are updated periodically. With the goal of minimizing energy consumption per time period, the optimization variables are optimized using the updated constraint boundaries. During the optimization process, the energy consumption prediction values ​​fed back in real time by the target total energy consumption prediction model are called. The optimization process is terminated when the number of iterations per single step is greater than or equal to a set number, or when the rate of change of energy consumption after a set number of consecutive iterations is less than or equal to a set rate of change. The optimization results for each optimization step are extracted, and the time-series set values ​​for each time period within the future set optimization cycle are integrated according to the time series. The time-series set values ​​include the specific values ​​of the optimization variables. Linear change processing is performed on the time-series set values, and outliers exceeding the rated operating range of the air conditioning equipment are removed and replaced with safe and feasible values. The target set value of the air conditioning system is then output.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that can be loaded by a processor and executed as a global optimization method for an air conditioning system as described in any one of claims 1 to 8.

Images