A Look-Ahead Optimization Control Method for Air Conditioning Systems Based on Rolling Prediction and Global Coordination

Abstract

This application belongs to the field of intelligent building and HVAC control technology, specifically disclosing a forward-looking optimization control method for air conditioning systems based on rolling prediction and global collaboration. The method includes: preprocessing and feature construction of air conditioning system data to obtain a historical data window; inputting this window into a Transformer prediction model to obtain a future prediction sequence; outputting a set of globally optimal setpoints through an optimization strategy generator; the optimization strategy generator first calculates the initial setpoints based on the prediction sequence using a set of empirical formulas, and then iteratively solves the system's total energy consumption model based on this initial setpoint, incorporating the physical coupling relationships between devices, to obtain the globally optimal setpoints that minimize the system's total energy consumption; the globally optimal setpoints are then sent to the device controllers for execution, and the Transformer prediction model is updated based on actual operating data. This application can solve the problem of large inertial lag in air conditioning systems and achieve accurate and reliable forward-looking energy efficiency optimization.

Description

Technical Field

[0001] This application belongs to the field of intelligent building and HVAC control technology, and more specifically, relates to a forward-looking optimization control method for air conditioning systems based on rolling prediction and global collaboration. Background Technology

[0002] Modern central air conditioning system control primarily relies on sensor-based feedback control strategies, such as adjusting the main unit's outlet water temperature based on the chilled water return temperature, or adjusting the pump frequency based on the most unfavorable terminal pressure difference. These methods are essentially passive response control, which has several drawbacks. First, air conditioning water systems are systems with high inertia and significant time lag. There is a substantial time delay between the issuance of control commands and the system's response and feedback, causing control to always lag behind actual load changes, leading to system oscillations and over-adjustment, impairing comfort and wasting energy. Second, traditional feedback control cannot predict future cooling / heating demands and weather changes, making proactive pre-adjustment impossible. This forces high-energy-consuming equipment to operate at high loads or frequently start and stop during peak load periods. Third, existing strategies often focus on independent control of single equipment or single loops, lacking global collaborative optimization of the cooling / heat source, distribution system, and terminals, resulting in suboptimal overall system energy efficiency. Furthermore, while some research has introduced load forecasting, it often employs traditional regression models or simple neural networks, which have limited ability to capture the complex nonlinear and long-term dependencies between weather, time, and building load, and the forecast accuracy is insufficient to support highly reliable optimized control.

[0003] Therefore, how to effectively overcome the large inertia lag problem of air conditioning systems and achieve accurate, reliable, and forward-looking energy efficiency optimization is an urgent problem to be solved. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of this application is to provide a forward-looking optimization control method for air conditioning systems based on rolling prediction and global coordination, which can solve the problem of large inertial lag in air conditioning systems and achieve accurate and reliable forward-looking energy efficiency optimization.

[0005] To achieve the above objectives, in a first aspect, this application provides a look-ahead optimization control method for an air conditioning system based on rolling prediction and global coordination, comprising the following steps:

[0006] S10 collects air conditioning system operation data, environmental data, and building operation data, and preprocesses and constructs features on the collected data to obtain a historical data window;

[0007] S20, input the historical data window into the pre-trained Transformer prediction model to obtain the prediction sequence for the next M time steps. The prediction sequence includes the total building cooling load, outdoor dry-bulb temperature and outdoor wet-bulb temperature.

[0008] S30, based on the predicted sequence, an optimization strategy generator calculates and outputs a set of globally optimal setpoints. The globally optimal setpoints include at least the chiller outlet water temperature setpoint, the cooling tower proximity setpoint, the chilled water system pressure difference setpoint, and the cooling water temperature difference setpoint. The optimization strategy generator first calculates the initial setpoints based on the predicted sequence using a set of empirical formulas. Then, starting from the initial setpoints, it iteratively solves the system total energy consumption model that incorporates the physical coupling relationships between devices to obtain the globally optimal setpoints that minimize the system's total energy consumption.

[0009] S40, the globally optimal setting value is sent to the corresponding device controller for execution, and the Transformer prediction model is updated based on the actual operating data.

[0010] The prospective optimization control method for air conditioning systems based on rolling prediction and global coordination provided in this application has the following effects: This method transforms traditional passive feedback control into proactive prospective optimization control by integrating high-precision rolling prediction with physical coupling-based global coordination optimization. Utilizing the high-precision prediction sequence of cooling load and meteorological parameters from the Transformer model as feedforward, the control system can issue control commands ahead of the large inertia of the air conditioning water system, thus fundamentally overcoming the problems of oscillation, overshoot, and energy waste caused by response lag. Simultaneously, by constructing and optimizing a global energy efficiency function that integrates the dynamic correlation between the cold source side and key equipment in the distribution system, joint decision-making on all key setpoints can break down control silos between equipment, automatically finding the operating point with the highest overall system energy efficiency, avoiding the "one-sided gain" phenomenon caused by local optimization, and achieving systemic energy saving. Furthermore, combined with an online model update mechanism, it ensures that the system can maintain long-term reliability of prediction and control even when facing changes in building usage patterns. Therefore, this method can achieve a fundamental shift from "passive response" to "proactive pre-adjustment," effectively improving the operating energy efficiency, stability, and adaptability of the air conditioning system.

[0011] As a further preferred embodiment, in step S20, the Transformer prediction model is a deep learning model based on an encoder structure, comprising N layers of stacked encoders, each layer containing a multi-head self-attention module and a feedforward neural network; the input feature sequence of the Transformer prediction model is a fixed-length historical data window, and the output sequence is the prediction sequence for a specific future time period.

[0012] As a further preferred embodiment, the Transformer prediction model is trained in the following manner:

[0013] During training, the load change prediction output by the model is clipped. The clipping formula is: ΔQ_pred'=clip(ΔQ_pred,-0.3σ,0.3σ), where ΔQ_pred is the original load change prediction value output by the model, ΔQ_pred' is the clipped load change prediction value, and σ is the standard deviation of the historical load change.

[0014] During training, the gradient of the loss function is clipped using the L2 norm. The specific steps include: calculating the L2 norm g_norm of the gradient vector; if g_norm ≥ Th, then scaling the gradient vector to (Th / g_norm) times; if g_norm < Th, then keeping the gradient unchanged; where Th is a preset gradient clipping threshold.

[0015] An early stopping mechanism is adopted, which terminates training when the validation set loss does not decrease within 20 consecutive training rounds.

[0016] As a further preferred embodiment, in step S30, the initial set value is calculated based on the predicted sequence using a set of empirical formulas, specifically including:

[0017] The initial value T_chw_set_init of the chiller unit outlet water temperature setpoint is calculated using the following formula:

[0018] T_chw_set_init = MEDIAN(a1 × (T_ldh - (Q_pred / Q_max) × ΔT_ld), T_ldh, T_ldg);

[0019] The initial value ΔT_approach_set_init for calculating the cooling tower approach setpoint is calculated using the following formula:

[0020] ΔT_approach_set_init = max(ΔT_approach_min, b1 × ln(T_out_pred) +b2 × ln²(Tw_out_pred) + b3);

[0021] The initial value ΔP_set_init of the differential pressure setpoint of the chilled water system is calculated using the following formula:

[0022] ΔP_set_init = min(0.5 × P_pump_rated, c1 × ((Q_pred / Q_max)² ×P_pump_rated / 1.2));

[0023] The initial value ΔT_set_init of the cooling water temperature difference setpoint is calculated using the following formula:

[0024] ΔT_set_init = MEDIAN(ΔT_set_min, d1 × (Q_pred / Q_max) × ΔT_cw_design, ΔT_cw_design);

[0025] Wherein, Q_pred is the predicted total building cooling load; Q_max is the maximum system load; ΔT_ld is the allowable outlet water temperature adjustment range under the chiller unit's design conditions; T_ldh is the upper limit of the outlet water temperature under the chiller unit's design conditions; T_ldg is the lower limit of the outlet water temperature under the chiller unit's design conditions; ΔT_approach_min is the minimum allowable value for cooling tower approach; T_out_pred is the predicted outdoor dry-bulb temperature; Tw_out_pred is the predicted outdoor wet-bulb temperature; P_pump_rated is the rated power of the water pump; ΔT_set_min is the minimum allowable value for cooling water temperature difference; ΔT_cw_design is the supply and return water temperature difference under the cooling water system's design conditions; a1, b1, b2, b3, c1, and d1 are empirical coefficients determined through historical data, where b3 is in °C.

[0026] As a further preferred option, the empirical coefficients a1, b1, b2, b3, c1, and d1 are determined by regression analysis or offline optimization of historical operating data of a specific air conditioning system.

[0027] As a further preferred embodiment, in step S10, the data preprocessing is as follows: missing values ​​are filled by interpolation, outliers are identified and corrected by statistical methods, and all feature data are normalized.

[0028] The feature construction refers to constructing features with physical or statistical significance, including historical load moving average features, future forecast values ​​of outdoor temperature and humidity features, population density prediction features based on schedule and historical data, and cumulative solar radiation features.

[0029] As a further preferred embodiment, steps S10 to S40 are executed in a rolling manner with a fixed control cycle. At the beginning of each control cycle, the latest historical data window is input into the pre-trained Transformer prediction model to obtain the prediction sequence for the next M time steps. In the next control cycle, the historical data window slides forward and the prediction is performed again to achieve rolling updates.

[0030] As a further preferred embodiment, in step S30, the total energy consumption model of the system is a global energy efficiency function P_sys(u), P_sys(u) = P_chiller(u) + P_pump(u) + P_fan(u), where u is the global optimization setpoint vector, P_chiller(u) is the energy consumption sub-function of the chiller unit, P_pump(u) is the total energy consumption sub-function of the water pump, and P_fan(u) is the energy consumption sub-function of the cooling tower fan;

[0031] The chiller unit energy consumption sub-function is P_chiller(u) = f1(Q_pred, T_chw_set, T_cond(u)); where Q_pred is the predicted total building cooling load, T_chw_set is the chiller unit outlet water temperature setpoint, T_cond(u) = g(ΔT_approach_set, Tw_out_pred, G_cw), ΔT_approach_set is the cooling tower approach setpoint, Tw_out_pred is the predicted outdoor wet-bulb temperature, and G_cw is the cooling water flow rate;

[0032] The total energy consumption subfunction of the water pump is P_pump(u) = P_chwp(u) + P_cwp(u); where P_chwp(u) = f2(ΔP_set, G_chw(Q_pred)), ΔP_set is the setpoint for the most unfavorable terminal pressure difference of the chilled water system, and G_chw(Q_pred) is the chilled water flow rate related to the predicted total building cooling load Q_pred; P_cwp(u) = f3(H_cw(ΔT_set), G_cw(Q_pred, ΔT_set)), ΔT_set is the setpoint for the cooling water temperature difference, H_cw(ΔT_set) is the cooling water pump head related to ΔT_set, and G_cw(Q_pred, ΔT_set) is the cooling water flow rate related to both Q_pred and ΔT_set;

[0033] The cooling tower fan energy consumption sub-function P_fan(u) = f4(ΔT_approach_set, G_cw(Q_pred,ΔT_set)); where f1, f2, f3, f4, and g are preset functional relationships characterizing the relationship between equipment performance and variables.

[0034] As a further preferred embodiment, in step S40, the Transformer prediction model is updated based on the actual operating data. Specifically, the deviation between the predicted load and the actual load is continuously monitored. When the prediction error exceeds a preset threshold for several consecutive days, or when a significant change in the building usage pattern is detected, the model update process is triggered. The Transformer prediction model is then fine-tuned online or periodically retrained using newly accumulated data.

[0035] Secondly, this application provides a global optimization control system for an air conditioning system, comprising the steps of implementing the method described in any one of the above, including:

[0036] Sensor networks are used to collect operating data of air conditioning systems, environmental data, and building operation data.

[0037] A data acquisition and storage unit, connected to the sensor network, is used to preprocess and construct features from the acquired data to obtain a historical data window;

[0038] The Transformer prediction model module, connected to the data acquisition and storage unit, is used to output a prediction sequence for the next M time steps based on the historical data window. The prediction sequence includes the building's total cooling load, outdoor dry-bulb temperature, and outdoor wet-bulb temperature.

[0039] A global optimization strategy generator, connected to the Transformer prediction model module, is used to calculate and output a set of globally optimal setpoints based on the prediction sequence. The globally optimal setpoints include at least the chiller outlet water temperature setpoint, the cooling tower approximation setpoint, the chilled water system pressure difference setpoint, and the cooling water temperature difference setpoint. The global optimization strategy generator is configured to first calculate the initial setpoints based on the prediction sequence using a set of empirical formulas, and then, starting from the initial setpoints, iteratively solve the system total energy consumption model that incorporates the physical coupling relationships between devices to obtain the globally optimal setpoints that minimize the system's total energy consumption.

[0040] The device controller is connected to the global optimization strategy generator and is used to receive and execute the global optimal setting value.

[0041] The beneficial effects of this application are as follows:

[0042] (1) Significant energy saving and consumption reduction (10%-20% or more): Through the coordinated optimization of multiple links such as "peak shaving and valley filling", increasing the average operating temperature of the host, optimizing the cooling tower proximity, and reducing the unnecessary head of the water pump, systematic energy saving is achieved.

[0043] (2) Look-ahead control overcomes system time delay: High-precision prediction is used for feedforward to compensate for the large inertia and large lag characteristics of the system in advance, so that the cooling capacity and demand are accurately matched in time, thereby fundamentally avoiding the inherent delay, oscillation and over-adjustment problems of traditional feedback control, and significantly improving the stability and comfort of indoor temperature and humidity environment.

[0044] (3) Achieve true global collaborative optimization: break down the control islands between equipment, optimize the key parameters of the cold source side and the transmission and distribution system in a coordinated manner, and pursue the optimal overall energy efficiency of the system.

[0045] (4) High prediction accuracy and strong adaptability: It adopts advanced deep learning models such as Transformer, which can accurately capture the complex nonlinear characteristics of load changes; combined with online learning mechanism, it can adapt to changes in building characteristics and usage patterns.

[0046] (5) Extend equipment life: Avoid operating equipment under extreme loads and frequent start-stop, and reduce mechanical and electrical wear.

[0047] (6) Strong engineering practicality: The system architecture is clear and can be seamlessly integrated with the existing building automation system (BAS). It can be achieved by adding edge computing devices to deploy algorithms, and is easy to promote.

[0048] (7) Achieve deep collaborative optimization based on physical coupling: By constructing and solving the system energy efficiency model that integrates the dynamic correlation of chiller, cooling tower and water pump online, the control island between equipment is broken through, and the optimization algorithm can automatically find the operating point with the highest global energy efficiency, avoiding the drawback of local optimization of "one gaining and the other losing", and further tapping the potential of system energy saving (it is expected to improve by 3%-8% on the basis of previous energy saving). Attached Figure Description

[0049] Figure 1 This is a flowchart of the look-ahead optimization control method for air conditioning systems based on rolling prediction and global coordination provided in this application;

[0050] Figure 2 This is a flowchart illustrating the overall architecture and closed-loop control of the method provided in the embodiments of this application.

[0051] Figure 3 This is a schematic diagram of the training and rolling prediction process of the Transformer heat load prediction model provided in the embodiments of this application;

[0052] Figure 4 This is a flowchart of the global optimization strategy generator provided in the embodiments of this application. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0054] like Figure 1 As shown, this application provides a look-ahead optimization control method for an air conditioning system based on rolling prediction and global coordination, including steps S10 to S40, which are detailed below:

[0055] Step S10: Collect air conditioning system operation data, environmental data, and building operation data, and preprocess and construct features for the collected data to obtain a historical data window.

[0056] This step allows for the systematic acquisition and clear reflection of multi-source heterogeneous data on system status and external environment, providing high-quality and standardized input features for subsequent model construction and laying a reliable data foundation for accurate prediction.

[0057] Step S20: Input the historical data window into the pre-trained Transformer prediction model to obtain the prediction sequence for the next M time steps. The prediction sequence includes the total building cooling load, outdoor dry-bulb temperature and outdoor wet-bulb temperature.

[0058] This step leverages the Transformer model's powerful ability to capture temporal features and model long-term dependencies, enabling the rolling output of high-precision predictions of cooling load and key meteorological parameters for the next short period, providing crucial feedforward input information for forward-looking control.

[0059] Step S30: Based on the predicted sequence, an optimization strategy generator calculates and outputs a set of globally optimal setpoints. The globally optimal setpoints include at least the chiller outlet water temperature setpoint, the cooling tower proximity setpoint, the chilled water system pressure difference setpoint, and the cooling water temperature difference setpoint. The optimization strategy generator first calculates the initial setpoints based on the predicted sequence using a set of empirical formulas. Then, starting from the initial setpoints, it iteratively solves the system total energy consumption model that incorporates the physical coupling relationships between devices to obtain the globally optimal setpoints that minimize the system's total energy consumption.

[0060] This step transforms predictive information into executable collaborative control commands. It employs a two-stage strategy, from rapid initial solution to deep global optimization, ensuring efficient solution of the global collaborative operating point that minimizes the total energy consumption of the entire air conditioning system, while considering the dynamic physical coupling between equipment such as chillers, pumps, and cooling towers. This breaks down control silos between devices and achieves optimal energy efficiency at the system level.

[0061] Step S40: The global optimal setting value is sent to the corresponding device controller for execution, and the Transformer prediction model is updated based on the actual operating data.

[0062] This step allows the optimized, forward-looking control setpoints to be issued to the equipment layer in advance, enabling control actions to anticipate the system's large inertia and fundamentally compensate for phase lag. Simultaneously, through an online model update mechanism based on operational feedback, the predictive model can continuously adapt to changes in building characteristics or usage patterns, maintaining predictive accuracy and system reliability for long-term applications.

[0063] The prospective optimization control method for air conditioning systems based on rolling prediction and global coordination provided in this application has the following effects: This method transforms traditional passive feedback control into proactive prospective optimization control by integrating high-precision rolling prediction with physical coupling-based global coordination optimization. Utilizing the high-precision prediction sequence of cooling load and meteorological parameters from the Transformer model as feedforward, the control system can issue control commands ahead of the large inertia of the air conditioning water system, thus fundamentally overcoming the problems of oscillation, overshoot, and energy waste caused by response lag. Simultaneously, by constructing and optimizing a global energy efficiency function that integrates the dynamic correlation between the cold source side and key equipment in the distribution system, joint decision-making on all key setpoints can break down control silos between equipment, automatically finding the operating point with the highest overall system energy efficiency, avoiding the "one-sided gain" phenomenon caused by local optimization, and achieving systemic energy saving. Furthermore, combined with an online model update mechanism, it ensures that the system can maintain long-term reliability of prediction and control even when facing changes in building usage patterns. Therefore, this method can achieve a fundamental shift from "passive response" to "proactive pre-adjustment," effectively improving the operating energy efficiency, stability, and adaptability of the air conditioning system.

[0064] In one embodiment, the technical solution to achieve the above objective can be as follows: This embodiment aims to provide a forward-looking control method capable of accurately predicting future short-term building heat load and key meteorological parameters, and generating globally optimal control setpoints accordingly. By transforming "passive response" control into "active pre-adjustment" control, the system time lag problem is fundamentally overcome, and a significant improvement in the energy efficiency and stability of the air conditioning system is achieved.

[0065] In view of the shortcomings of the prior art pointed out in the background section, the solution of this embodiment is as follows:

[0066] (1) To address the problems of insufficient accuracy and single dimension of the prediction model: This embodiment adopts a deep learning model based on Transformer, which can simultaneously capture the complex nonlinear relationship and long-term dependence between multiple dimensions of features such as meteorology, time, and historical load, and achieve hourly accurate prediction through a rolling prediction mechanism.

[0067] (2) To address the issues of lack of equipment synergy and local optimization: This embodiment constructs a global energy efficiency function P_sys(u) that integrates the physical coupling relationship between the chiller, water pump, and cooling tower. The set values ​​of each equipment are used as joint decision variables for optimization, which in principle achieves the optimal overall energy efficiency of the system and avoids the "one gains at the expense of the other".

[0068] (3) To address the issues of lack of foresight and delayed response: This embodiment uses the aforementioned high-precision prediction sequence as the feedforward quantity to construct a forward-looking control architecture of "prediction-optimization-feedforward", which enables the control command to be issued ahead of the system inertia, thereby compensating for the phase lag from the root.

[0069] like Figure 2 As shown, the overall steps are as follows:

[0070] Multi-source data acquisition and feature engineering: Real-time acquisition of historical and current system operation data and environmental data, followed by cleaning, normalization, and feature construction.

[0071] Heat load prediction model construction and rolling prediction: Construct a prediction model based on the Transformer deep learning architecture, and use its powerful ability to capture time series features to make rolling predictions of the building's hourly cooling load and outdoor dry / wet bulb temperature for a specific time window in the future.

[0072] Global optimization control strategy generation: Based on the predicted sequence, with the goal of minimizing the total energy consumption (P_sys) of the air conditioning system, and combined with the equipment performance model, the optimal operating setpoints of the chiller, cooling tower, chilled water pump and cooling water pump are solved collaboratively through optimization algorithms.

[0073] Proactive control command issuance and execution: The optimized setpoint sequence is issued to each device controller in advance, and combined with local feedback control to form a composite control to drive the actuator.

[0074] Feedback and online model self-updating: Based on the deviation between actual operating data and prediction results, the prediction model is updated through online fine-tuning or periodic retraining, enabling the system to have long-term adaptive capabilities.

[0075] The beneficial effects of this embodiment are:

[0076] (1) Significant energy saving and consumption reduction (10%-20% or more): Through the coordinated optimization of multiple links such as "peak shaving and valley filling", increasing the average operating temperature of the host, optimizing the cooling tower proximity, and reducing the unnecessary head of the water pump, systematic energy saving is achieved.

[0077] (2) Look-ahead control overcomes system time delay: High-precision prediction is used for feedforward to compensate for the large inertia and large lag characteristics of the system in advance, so that the cooling capacity and demand are accurately matched in time, thereby fundamentally avoiding the inherent delay, oscillation and over-adjustment problems of traditional feedback control, and significantly improving the stability and comfort of indoor temperature and humidity environment.

[0078] (3) Achieve true global collaborative optimization: break down the control islands between equipment, optimize the key parameters of the cold source side and the transmission and distribution system in a coordinated manner, and pursue the optimal overall energy efficiency of the system.

[0079] (4) High prediction accuracy and strong adaptability: It adopts advanced deep learning models such as Transformer, which can accurately capture the complex nonlinear characteristics of load changes; combined with online learning mechanism, it can adapt to changes in building characteristics and usage patterns.

[0080] (5) Extend equipment life: Avoid operating equipment under extreme loads and frequent start-stop, and reduce mechanical and electrical wear.

[0081] (6) Strong engineering practicality: The system architecture is clear and can be seamlessly integrated with the existing building automation system (BAS). It can be achieved by adding edge computing devices to deploy algorithms, and is easy to promote.

[0082] (7) Achieve deep collaborative optimization based on physical coupling: By constructing and solving the system energy efficiency model that integrates the dynamic correlation of chiller, cooling tower and water pump online, the control island between equipment is broken through, and the optimization algorithm can automatically find the operating point with the highest global energy efficiency, avoiding the drawback of local optimization of "one gaining and the other losing", and further tapping the potential of system energy saving (it is expected to improve by 3%-8% on the basis of previous energy saving).

[0083] The following is a specific implementation example of this application:

[0084] 1.1 System Hardware Configuration:

[0085] (1) Data Acquisition Layer:

[0086] Environmental sensors: outdoor dry / wet bulb temperature sensor, solar radiation sensor.

[0087] Indoor sensors: temperature and humidity sensors, CO2 sensors, and human presence sensors.

[0088] System sensors: chilled / cooling water supply and return temperature sensors, flow meters, differential pressure sensors for key branches, and power consumption meters for chiller / pump / cooling tower fans.

[0089] Other data sources: smart meters, building management system (BMS) schedules, access control / WiFi data (for population density estimation).

[0090] (2) Edge / cloud computing layer: industrial edge servers or cloud platforms that deploy prediction and optimization algorithms.

[0091] (3) Execution control layer: chiller unit controller, water pump frequency converter, cooling tower fan frequency converter, electric regulating valve, damper actuator, etc.

[0092] 1.2 Data Acquisition and Feature Engineering:

[0093] Data types: The collected historical and real-time data include time-series load values, outdoor meteorological parameters, indoor environmental parameters, equipment operating status and energy consumption, date type (weekday / holiday / special day), building schedule, etc.

[0094] Data preprocessing: missing values ​​are filled using interpolation, outliers are identified and corrected using statistical methods (such as the 3σ principle), and all feature data are normalized.

[0095] Feature construction: Construct features with physical or statistical significance, such as: historical load moving average, 24-hour forecast values ​​of outdoor temperature and humidity, population density prediction features based on schedule and historical data, and cumulative solar radiation.

[0096] 1.3 Construction and training of load forecasting models, such as Figure 3 As shown:

[0097] Model Selection: The Transformer deep learning model is adopted as the core predictor. Reason: Its multi-head self-attention mechanism can effectively capture the complex long-term dependencies between load and multi-dimensional influencing factors (such as temperature, humidity, historical load, and time context), overcoming the gradient vanishing problem that may exist in traditional recurrent neural networks (RNN / LSTM) during long-sequence training. It is particularly suitable for processing heat load time-series data with strong periodicity and nonlinearity.

[0098] In a preferred embodiment, the Transformer prediction model employs an encoder structure comprising N=3 stacked encoder layers. Each layer includes a multi-head self-attention module (h=4 heads, feature dimension d_model=64) and a feedforward neural network (hidden layer dimension d_ff=256). The input feature sequence length is 96 (corresponding to a 24-hour, 15-minute sampling interval), and the output sequence length is 24 (corresponding to a future 6-hour, 15-minute sampling interval). The activation function is ReLU, and the output layer uses the Sigmoid function to constrain the predicted value range to within [0, Q_max].

[0099] Model input and output:

[0100] Input (X): A fixed-length historical data window (e.g., the past 24 hours, with a sampling interval of 15 minutes) containing multidimensional feature vectors.

[0101] Output (Y): A predicted sequence of hourly total building cooling load (Q_pred), outdoor dry-bulb temperature (T_out_pred), and outdoor wet-bulb temperature (Tw_out_pred) for a specific future period (e.g., 2-6 hours from now).

[0102] 1.3.1 Model Training Optimization Strategy (Refinement of Invention Points):

[0103] Model architecture: The Transformer model adopts an encoder structure, and its output layer uses a bounded activation function to constrain the range of predicted values ​​and prevent numerical drift in long-term rolling predictions.

[0104] Feature engineering: In addition to historical load sequences, input features also include periodic features derived from timestamps (such as hourly sine and cosine coding) and features reflecting the system's operating status (such as the duration of zero load and the rate of change between adjacent time periods).

[0105] Training stability:

[0106] (1) The load change output of the model is clipped, ΔQ_pred'=clip(ΔQ_pred,-0.3σ,0.3σ), where σ is the standard deviation of the historical load change. This constraint is based on the physical law of short-term fluctuations in industrial load, avoids unreasonable abrupt predictions from the model, and enhances the smoothness and rationality of the prediction results.

[0107] (2) During training, the gradient of the loss function is clipped using the L2 norm to prevent gradient explosion when there are few samples. In a preferred embodiment, the gradient clipping steps are as follows:

[0108] 1. Calculate the gradient of the loss function for the current batch. L2 norm: g_norm= ;

[0109] 2. If g_norm ≥ 0.5, then scale the gradient: =(0.5 / g_norm)× ;

[0110] 3. If g_norm < 0.5, then keep the gradient unchanged: = ;

[0111] 4. Use the clipped gradient Update the model parameters.

[0112] (3) Monitor the validation set loss. If the validation set loss does not decrease for 20 consecutive training epochs, terminate the training early. This prevents the model from overfitting on limited data and ensures its generalization ability.

[0113] (4) Fix the random seed to ensure that the model training process is stable and the results can be completely reproduced.

[0114] Physical constraints: In the forward calculation of rolling forecast, physical constraints are applied to the forecast output at each step (such as ensuring that the load value is non-negative) so that the forecast sequence conforms to the actual operating law of the air conditioning system.

[0115] This embodiment, through the above design, effectively improves the robustness and reliability of the prediction model in industrial scenarios, and provides high-quality prediction input [Q_pred,T_out_pred,Tw_out_pred] for subsequent global optimization control.

[0116] 1.4 Generation of global optimization strategies based on joint solution of coupled models, such as Figure 4 As shown:

[0117] The core innovation of the global optimization control strategy generation in this embodiment lies in the fundamental innovation at both the control architecture and optimization paradigm levels.

[0118] First, in terms of control architecture, a transition was achieved from the traditional passive feedback control structure u(t)=f(e(t)) (based on the current or past error e(t)) to a forward-looking feedforward-feedback composite control structure u(t)=g(Φ(t+Δt)). Here, Φ(t+Δt) is the future time t+Δt heat load and weather forecast sequence [Q_pred,T_out_pred,Tw_out_pred] generated by the Transformer model. This architecture uses the high-confidence forecast sequence as a feedforward quantity, enabling the control system to make decisions ahead of the system's large inertia, thus compensating for phase lag in principle.

[0119] Secondly, in terms of optimization paradigm, based on this prediction sequence, the optimization object is upgraded from adjusting the independent parameters (single point) of multiple devices to jointly solving the global energy efficiency function P_sys(u) of the system, in order to pursue the overall optimum.

[0120] The global optimization strategy generator can adopt one of the following two implementation modes depending on the system's different requirements for real-time performance and optimization depth:

[0121] Mode 1 (Fast Optimization Mode): To meet scenarios with extremely high real-time requirements, the optimization strategy generator directly calculates and outputs global setpoints based on the predicted sequence using a set of pre-defined empirical formulas reflecting the relationship between device performance and load. This fast optimization mode can be used as an independent application of the first stage (fast initial value generation) in the aforementioned deep optimization mode, or as a simplified implementation method when computing resources are limited.

[0122] Mode 2 (Deep Optimization Mode): To achieve optimal global energy efficiency, the optimization strategy generator constructs a coupled system model and iteratively solves it. This mode will be explained in detail below.

[0123] 1.4.1 Rolling Forecasting Mechanism

[0124] The system operates on a fixed control cycle (e.g., 15 minutes). At the beginning of each control cycle, the latest historical data window is input into the trained Transformer model to obtain the predicted sequence [Q_pred, T_out_pred, Tw_out_pred] for the next M time steps. In the next cycle, the window slides forward, and predictions are re-established, achieving a rolling update.

[0125] 1.4.2 Optimization Modeling from Single-Point Setting to Global Function Construction

[0126] Traditional methods typically set independent optimization targets (single-point) for each device, while this embodiment constructs a unified system-wide energy efficiency function and uses the settings of all devices as its input variables for joint optimization.

[0127] Global optimization objective function:

[0128] Define the total power consumption of the system over a future control period as the objective function P_sys, which is a coupling function of the power of all major energy-consuming devices in the system:

[0129] P_sys(u)=P_chiller(u)+P_pump(u)+P_fan(u)

[0130] Where u is the global setpoint vector to be optimized.

[0131] Global setpoint vector (decision variables):

[0132] u=[T_chw_set,ΔT_approach_set,ΔP_set,ΔT_set] T

[0133] This vector integrates key control parameters for both the cold source and the distribution system.

[0134] Construct a global energy efficiency function:

[0135] The global function P_sys(u) is constructed by integrating the following sub-models, which characterize the coupling relationships between decision variables and device energy consumption, as well as among device energy consumption itself:

[0136] The coupled chiller unit energy consumption sub-function is: P_chiller(u) = f1(Q_pred, T_chw_set, T_cond(u)). Here, the condensing temperature T_cond is itself a function of the cooling tower approach setting ΔT_approach_set, the predicted wet-bulb temperature Tw_out_pred, and the cooling water flow rate (related to ΔT_set).

[0137] Coupled pump energy consumption sub-function: P_pump(u)=P_chwp(u)+P_cwp(u).

[0138] P_chwp(u) = f2(ΔP_set, G_chw(Q_pred)); P_cwp(u) = f3(H_cw(ΔT_set), G_cw(Q_pred, ΔT_set)). The flow rate G is related to the load Q_pred and the temperature difference setting ΔT_set.

[0139] Coupled cooling tower fan energy consumption function: P_fan(u)=f4(ΔT_approach_set,G_cw(Q_pred,ΔT_set)).

[0140] System constraints include the cooling balance Q_chiller_cap(u)≥Q_pred and equipment operating boundaries, which together constitute the constraints of the optimization problem.

[0141] 1.4.3 Efficient Two-Stage Solution Strategy

[0142] To efficiently solve the complex global function optimization problem described above, this embodiment adopts a two-stage strategy, balancing solution quality and real-time performance.

[0143] Phase 1: Rapid initial value generation based on empirical rules (utilizing the single-point optimization idea)

[0144] To provide a high-quality initial search point for global optimization, a set of decoupled empirical formulas are first used to quickly calculate an initial value u_init for each decision variable. These formulas are derived from engineering practice and reflect the basic relationship between each variable and the predicted load, for example:

[0145] T_chw_set_init=MEDIAN(a1×(T_ldh- ×ΔT_ld,T_ldh,T_ldg), dynamically calculated based on Q_pred to prioritize the host's high-efficiency area.

[0146] ΔT_approach_set_init=max(ΔT_approach_min,b1×ln(T_out_pred)+b2×ln²(Tw_out_pred)+b3), initialize a reasonable cooling water approximation based on the dynamic prediction of T_out_pred and Tw_out_pred.

[0147] ΔP_set_init=min(0.5×P_pump_rated,c1×( ()²×P_pump_rated / 1.2), based on Q_pred dynamic calculation, initially matching the terminal differential pressure requirement.

[0148] ΔT_set_init=MEDIAN(ΔT_set_min,d1× ×ΔT_cw_design, ΔT_cw_design), based on Q_pred dynamic calculation, initially match the cooling water temperature difference requirement.

[0149] This stage utilizes the speed of "single-point optimization," but the result u_init does not take into account the coupling between devices and is not the final solution.

[0150] Phase 2: Iterative optimization and solution of the global function

[0151] Starting with u_init, numerical optimization algorithms such as sequential quadratic programming or interior point method are used to iteratively solve the constructed global energy efficiency function P_sys(u). This process automatically searches in the solution space while weighing the gains and losses such as "energy saved by increasing the host water temperature" and "potential increase in pump energy consumption to maintain flow rate", and finally finds the globally optimal setpoint vector u that minimizes P_sys.

[0152] 1.4.4 Comparison of Technical Effects

[0153] Through the process described above, from "single-point initial value" to "global function optimization," this embodiment effectively solves the problem of inter-device target conflict in traditional methods. The output u ensures that all devices operate at the cooperative working point with the highest overall system energy efficiency, achieving a leap from local optimum to global optimum.

[0154] 1.5 Control Execution and Adaptive Updates:

[0155] Control command issuance: The optimized setpoint sequence is sent to the corresponding local controller (DDC / PLC) via standard industrial communication protocols (such as BACnet IP, Modbus TCP). The local controller combines the new setpoint with feedforward compensation and real-time feedback signals to execute high-precision composite PID control.

[0156] Model self-update mechanism: The system continuously monitors the deviation between predicted and actual loads. When the prediction error exceeds a preset threshold for several consecutive days, or when a significant change in building usage patterns is detected (such as renovation or functional adjustments), the model update process is automatically triggered. Utilizing recently accumulated new data, the Transformer prediction model is fine-tuned online or periodically retrained to ensure the system maintains high prediction accuracy and adaptability.

[0157] The symbols used in this embodiment are explained in the following table:

[0158]

[0159] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A look-ahead optimization control method for an air conditioning system based on rolling prediction and global coordination, characterized in that, Includes the following steps: S10 collects air conditioning system operation data, environmental data, and building operation data, and preprocesses and constructs features on the collected data to obtain a historical data window; S20, input the historical data window into the pre-trained Transformer prediction model to obtain the prediction sequence for the next M time steps. The prediction sequence includes the total building cooling load, outdoor dry-bulb temperature and outdoor wet-bulb temperature. S30, based on the predicted sequence, an optimization strategy generator calculates and outputs a set of globally optimal setpoints. The globally optimal setpoints include at least the chiller outlet water temperature setpoint, the cooling tower proximity setpoint, the chilled water system pressure difference setpoint, and the cooling water temperature difference setpoint. The optimization strategy generator first calculates the initial setpoints based on the predicted sequence using a set of empirical formulas. Then, starting from the initial setpoints, it iteratively solves the system total energy consumption model that incorporates the physical coupling relationships between devices to obtain the globally optimal setpoints that minimize the system's total energy consumption. S40, the global optimal setting value is sent to the corresponding device controller for execution, and the Transformer prediction model is updated according to the actual operating data; In step S30, the initial set value is calculated based on the predicted sequence using a set of empirical formulas, specifically including: The initial value T_chw_set_init of the chiller unit outlet water temperature setpoint is calculated using the following formula: T_chw_set_init = MEDIAN(a1 × (T_ldh - (Q_pred / Q_max) × ΔT_ld), T_ldh, T_ldg); The initial value ΔT_approach_set_init for calculating the cooling tower approach setpoint is calculated using the following formula: ΔT_approach_set_init = max(ΔT_approach_min, b1 × ln(T_out_pred) + b2× ln²(Tw_out_pred) + b3); The initial value ΔP_set_init of the differential pressure setpoint of the chilled water system is calculated using the following formula: ΔP_set_init = min(0.5 × P_pump_rated, c1 × ((Q_pred / Q_max)² × P_pump_rated / 1.2)); The initial value ΔT_set_init of the cooling water temperature difference setpoint is calculated using the following formula: ΔT_set_init = MEDIAN(ΔT_set_min, d1 × (Q_pred / Q_max) × ΔT_cw_design, ΔT_cw_design); Wherein, Q_pred is the predicted total building cooling load; Q_max is the maximum system load; ΔT_ld is the allowable outlet water temperature adjustment range under the chiller unit's design conditions; T_ldh is the upper limit of the outlet water temperature under the chiller unit's design conditions; T_ldg is the lower limit of the outlet water temperature under the chiller unit's design conditions; ΔT_approach_min is the minimum allowable value for cooling tower approach; T_out_pred is the predicted outdoor dry-bulb temperature; Tw_out_pred is the predicted outdoor wet-bulb temperature; P_pump_rated is the rated power of the water pump; ΔT_set_min is the minimum allowable value for cooling water temperature difference; ΔT_cw_design is the supply and return water temperature difference under the cooling water system's design conditions; a1, b1, b2, b3, c1, and d1 are empirical coefficients determined through historical data, where b3 is in °C. In step S30, the total energy consumption model of the system is a global energy efficiency function P_sys(u), P_sys(u) = P_chiller(u) + P_pump(u) + P_fan(u), where u is the global optimization setpoint vector, P_chiller(u) is the energy consumption sub-function of the chiller unit, P_pump(u) is the total energy consumption sub-function of the water pump, and P_fan(u) is the energy consumption sub-function of the cooling tower fan. The chiller unit energy consumption sub-function is P_chiller(u) = f1(Q_pred, T_chw_set, T_cond(u)); where Q_pred is the predicted total building cooling load, T_chw_set is the chiller unit outlet water temperature setpoint, T_cond(u) = g(ΔT_approach_set, Tw_out_pred, G_cw), ΔT_approach_set is the cooling tower approach setpoint, Tw_out_pred is the predicted outdoor wet-bulb temperature, and G_cw is the cooling water flow rate; The total energy consumption subfunction of the water pump is P_pump(u) = P_chwp(u) + P_cwp(u); where, P_chwp(u) = f2(ΔP_set, G_chw(Q_pred)), ΔP_set is the setpoint for the most unfavorable terminal pressure difference of the chilled water system, and G_chw(Q_pred) is the chilled water flow rate related to the predicted total building cooling load Q_pred; P_cwp(u) = f3(H_cw(ΔT_set), G_cw(Q_pred, ΔT_set)), ΔT_set is the setpoint for the cooling water temperature difference, H_cw(ΔT_set) is the cooling water pump head related to ΔT_set, and G_cw(Q_pred, ΔT_set) is the cooling water flow rate related to both Q_pred and ΔT_set; The cooling tower fan energy consumption sub-function P_fan(u) = f4(ΔT_approach_set, G_cw(Q_pred, ΔT_set)); where f1, f2, f3, f4, and g are preset functional relationships characterizing the relationship between equipment performance and variables.

2. The look-ahead optimization control method for air conditioning systems based on rolling prediction and global coordination as described in claim 1, characterized in that, In step S20, the Transformer prediction model is a deep learning model based on an encoder structure, which includes N layers of encoder stacked together. Each layer includes a multi-head self-attention module and a feedforward neural network. The input feature sequence of the Transformer prediction model is a fixed-length historical data window, and the output sequence is the prediction sequence for a specific future time period.

3. The look-ahead optimization control method for air conditioning systems based on rolling prediction and global coordination as described in claim 1, characterized in that, The Transformer prediction model is trained in the following way: During training, the load change prediction output by the model is clipped. The clipping formula is: ΔQ_pred'=clip(ΔQ_pred,-0.3σ,0.3σ), where ΔQ_pred is the original load change prediction value output by the model, ΔQ_pred' is the clipped load change prediction value, and σ is the standard deviation of the historical load change. During training, the gradient of the loss function is clipped using the L2 norm. The specific steps include: calculating the L2 norm g_norm of the gradient vector; if g_norm ≥ Th, then scaling the gradient vector to (Th / g_norm) times; if g_norm < Th, then keeping the gradient unchanged; where Th is a preset gradient clipping threshold. An early stopping mechanism is adopted, which terminates training when the validation set loss does not decrease within 20 consecutive training rounds.

4. The look-ahead optimization control method for air conditioning systems based on rolling prediction and global coordination as described in claim 1, characterized in that, The empirical coefficients a1, b1, b2, b3, c1, and d1 are determined by regression analysis or offline optimization of historical operating data of a specific air conditioning system.

5. The look-ahead optimization control method for air conditioning systems based on rolling prediction and global coordination as described in claim 1, characterized in that, In step S10, the data preprocessing is as follows: missing values ​​are filled using interpolation, outliers are identified and corrected using statistical methods, and all feature data are normalized. The feature construction refers to constructing features with physical or statistical significance, including historical load moving average features, future forecast values ​​of outdoor temperature and humidity features, population density prediction features based on schedule and historical data, and cumulative solar radiation features.

6. The look-ahead optimization control method for air conditioning systems based on rolling prediction and global coordination as described in claim 1, characterized in that, Steps S10 to S40 are executed in a fixed control cycle. At the beginning of each control cycle, the latest historical data window is input into the pre-trained Transformer prediction model to obtain the prediction sequence for the next M time steps. In the next control cycle, the historical data window slides forward and the prediction is performed again to achieve rolling updates.

7. The look-ahead optimization control method for air conditioning systems based on rolling prediction and global coordination as described in claim 1, characterized in that, In step S40, the Transformer prediction model is updated based on the actual operating data. Specifically, the deviation between the predicted load and the actual load is continuously monitored. When the prediction error exceeds the preset threshold for several consecutive days, or when a significant change in the building usage pattern is detected, the model update process is triggered. The Transformer prediction model is then fine-tuned online or periodically retrained using the recently accumulated new data.

8. A global optimization control system for an air conditioning system, characterized in that, The steps for implementing the method according to any one of claims 1 to 7 include: Sensor networks are used to collect operating data of air conditioning systems, environmental data, and building operation data. A data acquisition and storage unit, connected to the sensor network, is used to preprocess and construct features from the acquired data to obtain a historical data window; The Transformer prediction model module, connected to the data acquisition and storage unit, is used to output a prediction sequence for the next M time steps based on the historical data window. The prediction sequence includes the building's total cooling load, outdoor dry-bulb temperature, and outdoor wet-bulb temperature. A global optimization strategy generator, connected to the Transformer prediction model module, is used to calculate and output a set of globally optimal setpoints based on the prediction sequence. The globally optimal setpoints include at least the chiller outlet water temperature setpoint, the cooling tower approximation setpoint, the chilled water system pressure difference setpoint, and the cooling water temperature difference setpoint. The global optimization strategy generator is configured to first calculate the initial setpoints based on the prediction sequence using a set of empirical formulas, and then, starting from the initial setpoints, iteratively solve the system total energy consumption model that incorporates the physical coupling relationships between devices to obtain the globally optimal setpoints that minimize the system's total energy consumption. The device controller is connected to the global optimization strategy generator and is used to receive and execute the global optimal setting value.

Images