A heating and refrigeration water collaborative control method and system based on a double-branch deep and shallow feature fusion neural network

Abstract

The application discloses a heating and refrigeration water cooperative control method and system based on a double-branch deep and shallow feature fusion neural network, and the method comprises the following steps: obtaining historical operation data of a heating and refrigeration water system and pre-processing the historical operation data to form standardized training data; training a double-branch deep and shallow feature fusion neural network model based on the standardized training data, wherein the model realizes cooperative prediction of equipment through a cold water host load prediction branch and a refrigeration pump frequency prediction branch in combination with a deep and shallow feature fusion mechanism; collecting environmental parameters in real time and inputting the environmental parameters into the trained model to output a cold water host load prediction value and a refrigeration pump frequency prediction value; and correcting the prediction values based on a physical constraint correction algorithm and outputting an optimized control instruction. The application solves the technical problems of insufficient equipment cooperation, insufficient feature extraction and poor control reliability caused by the lack of physical constraints in the prior art.

Description

Technical Field

[0001] This invention relates to the field of HVAC control technology, and in particular to a collaborative control method and system for HVAC chilled water based on a dual-branch deep and shallow feature fusion neural network. Background Technology

[0002] With increasingly stringent building energy efficiency requirements, energy consumption optimization of HVAC systems has become a core issue in the field of green building. As a key energy-consuming component of HVAC systems, the quality of the control strategy for chilled water systems directly impacts the overall system's energy efficiency. Traditional control methods, such as PID control, suffer from issues like response lag and over-reliance on expert experience. While existing single-device optimization strategies can improve local energy efficiency, they neglect the coupling relationship between the chiller and the chilled water pump, making it difficult to achieve system-level global optimization. Furthermore, purely data-driven neural network models often produce unreasonable results in practical applications due to a lack of consideration for physical constraints, posing operational safety risks.

[0003] Current chilled water system control technology faces the following challenges: First, insufficient coordination between equipment leads to local optima but overall inefficiency; second, traditional feature extraction methods struggle to capture the deep correlation between environmental parameters and equipment status; and third, there is a lack of mechanisms to integrate physical laws into data-driven models, affecting the reliability and safety of control. Therefore, there is an urgent need for a collaborative control method and system for HVAC chilled water systems that can achieve coordinated equipment optimization, deeply integrate environmental features and equipment status, and embed physical constraints. Summary of the Invention

[0004] The main objective of this invention is to propose a collaborative control method and system for HVAC chilled water based on a dual-branch deep and shallow feature fusion neural network, aiming to solve the technical problems of insufficient equipment collaboration, inadequate feature extraction, and poor control reliability caused by lack of physical constraints in the prior art.

[0005] To achieve the above objectives, the first aspect of this invention proposes a collaborative control method for HVAC chilled water based on a dual-branch deep and shallow feature fusion neural network, comprising:

[0006] Step S100: Obtain historical operating data of the HVAC chilled water system, preprocess the historical operating data to form standardized training data; the historical operating data includes environmental parameters and equipment operating parameters;

[0007] Step S200: Train the dual-branch deep and shallow feature fusion neural network model based on standardized training data until the model converges; the dual-branch deep and shallow feature fusion neural network model includes a chiller load prediction branch and a chilled pump frequency prediction branch, and realizes the joint prediction of chiller load and chilled pump frequency opening through feature fusion mechanism.

[0008] Step S300: Collect environmental parameters of the HVAC chilled water system in real time, preprocess them, and input them into the trained dual-branch deep and shallow feature fusion neural network model to output the chiller load prediction value and the chilled pump frequency prediction value.

[0009] Step S400: Based on the preset constraint correction algorithm, correct the predicted load value of the chiller and the predicted frequency value of the chilled water unit, and output the corrected control command.

[0010] Preferably, in step S100, the steps of acquiring historical operating data of the HVAC chilled water system and preprocessing the historical operating data to form standardized training data include:

[0011] Step S110: Collect historical operating data of the HVAC chilled water system. Environmental parameters include outdoor temperature, outdoor relative humidity, solar radiation intensity, number of people indoors and indoor terminal opening. Equipment operating parameters include chiller load rate and chilled water pump frequency.

[0012] Step S120: Denoise the collected historical operating data to remove outliers caused by sensor malfunctions and invalid data that are outside the physical reasonable range;

[0013] Step S130: Standardize the cleaned historical running data to form training dataset and test dataset.

[0014] Preferably, in step S200, the dual-branch deep and shallow feature fusion neural network model includes a chiller load prediction branch and a chilled pump frequency prediction branch. The steps for jointly predicting the chiller load and chilled pump frequency through the feature fusion mechanism include:

[0015] Step S210: Extract shallow environmental features based on standardized training data, and transform the shallow environmental features into deep environmental features through neural network transformation;

[0016] Step S220: Input the deep environmental characteristics into the chiller load prediction branch, and output the chiller load prediction value and deep chiller characteristics;

[0017] Step S230: Fuse the shallow environment features, deep environment features, and deep chiller features to generate fused features;

[0018] Step S240: Input the fused features into the cryogenic pump model and output the predicted cryogenic pump frequency.

[0019] Preferably, step S210 includes:

[0020] Step S211: Input the preprocessed environmental parameters as shallow environmental features, which include outdoor temperature, outdoor relative humidity, solar radiation intensity, number of people indoors, and indoor terminal opening degree.

[0021] Step S212: The shallow environmental features are transformed nonlinearly through the cold load common feature extraction module. The cold load common feature extraction module adopts a multi-layer fully connected neural network structure, which includes an input layer, at least one hidden layer and an output layer.

[0022] Step S213: Mind the correlation between environmental parameters and cooling load through the activation function of the hidden layer, and output deep environmental features.

[0023] Preferably, step S220 includes:

[0024] Step S221: Input the deep environmental features into the neural network of the chiller load prediction branch. The neural network includes an input layer, at least one hidden layer and an output layer.

[0025] Step S222: Through nonlinear transformation of the hidden layer, the deep environmental features are compressed and correlations are mined to generate intermediate features;

[0026] Step S223: The output layer uses an activation function to generate a predicted load value for the chiller unit and outputs the intermediate features as deep chiller unit features.

[0027] Preferably, step S230 includes:

[0028] Step S231: Map the shallow environmental features to the same dimension as the deep environmental features through the projection layer to achieve dimension matching;

[0029] Step S232: Using a residual connection mechanism, the mapped shallow environment features, deep environment features, and deep chiller unit features are superimposed and fused.

[0030] Step S233: Apply an activation function to the superimposed and fused features to perform a nonlinear transformation, generating fused features that include environmental details and device status information.

[0031] Preferably, step S240 includes:

[0032] Step S241: Input the fused features into the neural network of the cryopump frequency prediction branch, the neural network including an input layer, at least one hidden layer and an output layer;

[0033] Step S242: The fusion features are processed by nonlinear transformation of the hidden layer to capture the coupling relationship between environmental requirements and equipment status;

[0034] Step S243: The output layer uses an activation function to generate normalized predicted values ​​for the refrigeration pump frequency.

[0035] Preferably, step S400 includes:

[0036] Step S410: Based on the deviation between the supply and return water temperature difference and the ideal temperature difference, proportionally correct the predicted value of the chilled water pump frequency;

[0037] Step S420: Perform boundary constraint processing on the corrected chiller load prediction value and chilled pump frequency prediction value to ensure that they are within the safe range of equipment operation;

[0038] Step S430: Output the final control command that satisfies the physical constraints.

[0039] Preferably, in step S200, the step of training the dual-branch deep and shallow feature fusion neural network model based on standardized training data until the model converges includes:

[0040] Step S250: Based on standardized training data, perform the forward propagation process of the dual-branch deep and shallow feature fusion network model to obtain the joint predicted value of chiller load and chilled pump frequency, and construct a comprehensive loss function, which includes chiller load prediction loss, chilled pump frequency prediction loss and system energy consumption loss.

[0041] Step S260: Calculate the gradient of the comprehensive loss function with respect to the model parameters using the gradient descent algorithm, and optimize the parameters through backpropagation to minimize the comprehensive loss function;

[0042] Step S270: Iteratively execute the forward and backward propagation processes. When the loss of the comprehensive loss function satisfies the preset convergence condition, stop training and save the optimal model parameters.

[0043] A second aspect of this invention proposes a HVAC chilled water collaborative control system based on a dual-branch deep and shallow feature fusion neural network, comprising:

[0044] The data acquisition and preprocessing module is used to acquire historical operating data of the HVAC chilled water system, preprocess the historical operating data to form standardized training data; the historical operating data includes environmental parameters and equipment operating parameters.

[0045] The dual-branch model training module is used to train the dual-branch deep and shallow feature fusion neural network model based on standardized training data until the model converges. The dual-branch deep and shallow feature fusion neural network model includes a chiller load prediction branch and a chilled pump frequency prediction branch, which realizes the joint prediction of chiller load and chilled pump frequency opening through feature fusion mechanism.

[0046] The real-time predictive control module is used to collect environmental parameters of the HVAC chilled water system in real time, preprocess them, and then input them into a trained dual-branch deep and shallow feature fusion neural network model to output the predicted values ​​of chiller load and chilled water pump frequency.

[0047] The constraint correction module is used to correct the predicted load values ​​of the chiller and the predicted frequency values ​​of the chilled water unit based on a preset constraint correction algorithm, and output the corrected control commands.

[0048] This invention provides a collaborative control method and system for HVAC chilled water based on a dual-branch shallow and deep feature fusion neural network. The method achieves collaborative prediction between the chiller and chilled water pump through the dual-branch shallow and deep feature fusion neural network, improving the system's global optimization capability and control accuracy. It integrates shallow environmental features with deep equipment states through a feature fusion mechanism, enhancing the model's adaptability to complex operating conditions and predictive timeliness. A physical constraint correction module ensures the output conforms to equipment operating patterns, improving system safety and reliability. Neural network modeling reduces reliance on expert experience, improving intelligent control efficiency and practicality. Training based on historical data enables the model to adapt to changing operating conditions, improving energy-saving effects and generalization ability.

[0049] Furthermore, this invention improves the completeness of the training dataset by collecting multi-dimensional environmental and equipment parameters; improves the quality of the training dataset by cleaning abnormal data; enhances the model's feature representation ability by extracting shallow and deep environmental features; improves the control coordination between devices through dual-branch collaborative prediction; improves the accuracy of the chilled pump model's understanding of complex operating conditions through a deep and shallow feature fusion mechanism; improves the response speed to sudden operating conditions by retaining the original environmental parameters as shallow environmental features; improves the accuracy of cooling load prediction by mining deep correlations between parameters through a common feature extraction module; improves the accuracy of chiller unit load prediction by inputting deep environmental features into the load prediction branch; and improves the accuracy of chiller unit load prediction by outputting intermediate features as... The deep chiller host features are used to improve the information integrity of subsequent feature fusion; dimension matching is achieved through a projection layer to improve the compatibility of feature fusion; the integrity of feature information is improved through a residual connection mechanism; the representational ability of fused features is improved through nonlinear transformation of activation functions; the environmental adaptability of chiller pump control is improved by using the fused feature input chiller pump frequency prediction branch; control accuracy is improved by temperature difference deviation ratio correction; operational safety is improved by boundary constraint processing; system reliability is improved by physical constraint output; multi-objective optimization capability is improved by constructing a comprehensive loss function; training efficiency is improved by optimizing parameters through gradient descent algorithm; and model generalization ability is improved by stopping training through verification loss convergence conditions.

[0050] In summary, the present invention proposes a collaborative control method and system for HVAC chilled water based on a dual-branch deep and shallow feature fusion neural network, which solves the technical problems of insufficient equipment collaboration, inadequate feature extraction, and poor control reliability caused by lack of physical constraints in the prior art. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0052] Figure 1 A flowchart of a collaborative control method for HVAC chilled water based on a dual-branch deep and shallow feature fusion neural network provided in an embodiment of the present invention;

[0053] Figure 2 This is a flowchart of a method for acquiring and preprocessing multi-source historical running data according to an embodiment of the present invention;

[0054] Figure 3 A flowchart illustrating a method for jointly predicting chiller load and chilled water pump frequency using a feature fusion mechanism, as provided in an embodiment of the present invention.

[0055] Figure 4 A flowchart illustrating a method for extracting deep and shallow environmental features based on standardized training data, provided in an embodiment of the present invention;

[0056] Figure 5 A flowchart illustrating a method for generating chiller load prediction values ​​and deep chiller characteristics based on deep environmental characteristics, according to an embodiment of the present invention;

[0057] Figure 6 A flowchart illustrating a method for integrating shallow environment characteristics, deep environment characteristics, and deep chiller characteristics according to an embodiment of the present invention;

[0058] Figure 7 A flowchart of a method for generating predicted values ​​of refrigeration pump frequency based on fusion features according to an embodiment of the present invention;

[0059] Figure 8 The flowchart illustrates a method for correcting the predicted load value of a chiller and the predicted frequency value of a chilled water unit based on a preset constraint correction algorithm, according to an embodiment of the present invention.

[0060] Figure 9 A flowchart of a method for training a dual-branch deep and shallow feature fusion neural network model based on standardized training data, according to an embodiment of the present invention;

[0061] Figure 10 This is a schematic diagram of a HVAC chilled water collaborative control system based on a dual-branch deep and shallow feature fusion neural network provided in an embodiment of the present invention;

[0062] Figure 11 This is a schematic diagram of the structure of a dual-branch deep and shallow feature fusion model provided in an embodiment of the present invention.

[0063] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] It should be noted that if the embodiments of the present invention involve directional indicators, such as up, down, left, right, front, back, etc., the directional indicators are only used to explain the relative positional relationship and movement of the components in a specific posture. If the specific posture changes, the directional indicators will also change accordingly.

[0066] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the meaning of "and / or" throughout the text includes three parallel solutions; for example, "A and / or B" includes solution A, solution B, or a solution where both A and B are satisfied simultaneously. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0067] The main objective of this invention is to propose a collaborative control method and system for HVAC chilled water based on a dual-branch deep and shallow feature fusion neural network, aiming to solve the technical problems of insufficient equipment collaboration, inadequate feature extraction, and poor control reliability caused by lack of physical constraints in the prior art.

[0068] like Figures 1 to 11As shown, the first aspect of this invention proposes a collaborative control method for HVAC chilled water based on a dual-branch deep and shallow feature fusion neural network, comprising:

[0069] Step S100: Obtain historical operating data of the HVAC chilled water system, preprocess the historical operating data to form standardized training data; the historical operating data includes environmental parameters and equipment operating parameters;

[0070] Step S200: Train the dual-branch deep and shallow feature fusion neural network model based on standardized training data until the model converges; the dual-branch deep and shallow feature fusion neural network model includes a chiller load prediction branch and a chilled pump frequency prediction branch, and realizes the joint prediction of chiller load and chilled pump frequency opening through feature fusion mechanism.

[0071] Step S300: Collect environmental parameters of the HVAC chilled water system in real time, preprocess them, and input them into the trained dual-branch deep and shallow feature fusion neural network model to output the chiller load prediction value and the chilled pump frequency prediction value.

[0072] Step S400: Based on the preset constraint correction algorithm, correct the predicted load value of the chiller and the predicted frequency value of the chilled water unit, and output the corrected control command.

[0073] For details, see Figure 1In a specific embodiment of the present invention, historical operating data of the HVAC chilled water system is first collected through an HVAC monitoring system and an indoor and outdoor sensor network. This includes environmental parameters such as outdoor temperature, outdoor relative humidity, solar radiation intensity, number of people indoors, and indoor terminal opening, as well as equipment operating parameters such as chiller load rate and chilled water pump frequency. Then, this raw data is cleaned to remove outliers caused by sensor malfunctions or physical defects, and standardized to form training and testing datasets. Subsequently, a dual-branch deep and shallow feature fusion neural network model is trained using the standardized data. The core of this model consists of a chiller load prediction branch and a chilled water pump frequency prediction branch. The two branches are collaboratively optimized through a feature fusion mechanism. The training process iterates continuously until the model loss function converges, ensuring that the model can adapt to changes in different operating conditions. The training employs a gradient descent optimization algorithm to minimize the comprehensive loss function. The comprehensive loss function is set to balance prediction accuracy and system energy consumption. Training stops and the optimal parameters are saved when the verification loss meets the preset convergence condition. In the practical application phase, environmental parameters of the HVAC chilled water system are collected in real time. After the same preprocessing as in the training phase, these parameters are input into the trained model. The model sequentially undergoes feature extraction, bi-branch prediction, and fusion calculation, outputting predicted values ​​for the chiller load and chilled pump frequency. Finally, the predicted values ​​for the chiller load and chilled pump frequency are corrected based on physical constraints. The chilled pump frequency is fine-tuned proportionally by adjusting the deviation between the supply and return water temperature difference and the ideal temperature difference. Equipment operating boundary constraints are applied to the chiller load and chilled pump frequency to ensure that the output values ​​are within a safe range, thereby generating the final control command to drive the field equipment. This entire method combines data-driven approaches with physical principles to achieve adaptive and coordinated control of the chilled water system, dynamically optimizing energy consumption without relying on human experience.

[0074] Understandably, this embodiment achieves collaborative prediction between the chiller and the chilled water pump through a dual-branch deep and shallow feature fusion neural network, improving the system's global optimization capability and control accuracy; it integrates shallow environmental features with deep equipment states through a feature fusion mechanism, improving the model's adaptability to complex operating conditions and predictive timeliness; it ensures that the output conforms to the equipment's operating rules through a physical constraint correction module, improving system safety and reliability; it reduces reliance on expert experience through neural network modeling, improving intelligent control efficiency and practicality; and it enables the model to adapt to changing operating conditions based on historical data training, improving energy-saving effects and generalization ability.

[0075] Preferably, step S100 includes:

[0076] Step S110: Collect historical operating data of the HVAC chilled water system. Environmental parameters include outdoor temperature, outdoor relative humidity, solar radiation intensity, number of people indoors and indoor terminal opening. Equipment operating parameters include chiller load rate and chilled water pump frequency.

[0077] Step S120: Clean the collected historical operating data to remove outliers caused by sensor malfunctions and invalid data that are outside the physical reasonable range;

[0078] Step S130: Standardize the historical operating data after cleaning to form a training dataset and a test dataset, in which the chiller load rate is expressed as a percentage and the chilled pump frequency is normalized to the load value.

[0079] For details, see Figure 2 In a specific embodiment of the present invention, multi-source historical operating data is first collected from the actual operating HVAC chilled water system. This historical operating data must comprehensively cover typical operating conditions such as peak summer load periods, transitional seasons, and low winter load periods to ensure the model can adapt to various operating conditions throughout the year. The data sources are the HVAC monitoring system and indoor / outdoor sensor networks, ensuring the integrity and accuracy of the data through these two channels. Based on the operating characteristics and theory of the HVAC chilled water system, five types of key input data and two types of output data are selected to comprehensively capture the dynamic changes in building cooling load. The input data includes outdoor temperature. T out Outdoor relative humidity RH out Surface radiation intensity of buildings I rad Total number of people indoors N people Total opening of indoor terminals O terminal The output tag data includes the chiller load rate. y chiller and chilled pump frequency f pump Among them, outdoor temperature T out The unit is degrees Celsius (°C), and the value range is typically from -5 degrees Celsius in winter to 40 degrees Celsius in summer. In most parts of my country, summer outdoor temperatures are concentrated between 25 and 35 degrees Celsius. For every 1 degree Celsius increase in outdoor temperature, the cooling load increases by approximately 3% to 5%, indicating that the outdoor thermal level is the most significant factor affecting the heat transfer load of the building envelope; outdoor relative humidity... RH out This refers to the latent heat load of a building, expressed as a percentage (%), ranging from 30% to 90%. In summer, when the relative humidity is above 70%, the latent heat load can account for 30% to 40% of the total cooling load. For every 10% increase in outdoor humidity, the cooling load increases by approximately 5% to 8%, directly affecting the building's latent heat load; the building's exterior radiation intensity... I radThe unit is watts per square meter, ranging from 0 W / m² at night to 1000 W / m² on a sunny midday. In summer, the midday solar radiation intensity can reach 800 to 1000 W / m², at which point solar radiation heat gain accounts for 20% to 30% of the total cooling load. For every 100 W / m² increase in solar radiation intensity, the cooling load increases by approximately 2% to 3%, which is a significant component of the summer cooling load; total number of people indoors. N people The unit is people, reflecting the sensible and latent heat load generated by people indoors. Person density is usually expressed as people per square meter, ranging from 0 (empty) to 5. In densely populated areas such as conference rooms, each person generates approximately 100W to 150W of sensible heat and 50W to 100W of latent heat per hour. For every increase of 1 person per square meter in personnel density, the cooling load increases by approximately 10% to 15%. The total opening degree of indoor terminals... O terminal The unit is percentage (%), which is the average opening degree of all air conditioning terminal valves or dampers, ranging from 0% closed to 100% fully open. The larger the opening degree, the more cooling capacity the terminal needs and the higher the cooling load. For every 10% increase in the opening degree of the terminal equipment, the cooling load increases by approximately 8% to 10%, indirectly reflecting the current indoor cooling load level; the output label data includes the chiller load rate. y chiller and chilled pump frequency f pump Chiller load rate y chiller The unit is percentage (%), representing the ratio of the actual cooling capacity of the chiller to its rated cooling capacity; chilled water pump frequency. f pump The unit is Hertz (Hz), which needs to be normalized to a load value between 0 and 1 according to the pump's rated frequency. The conversion formula is as follows:

[0080]

[0081] In the formula, This represents the normalized refrigeration pump load value. ; This indicates the actual operating frequency of the chilled water pump. This indicates the minimum operating frequency of the refrigeration pump. This indicates the maximum operating frequency of the refrigeration pump.

[0082] After cleaning and preprocessing the collected raw data, removing numerical jumps caused by sensor malfunctions and outliers exceeding the physical reasonable range, the training dataset is finally formed. ,in m The number of samples in the training set, and the input vector. The output data is Similarly, the test dataset can be obtained. , nThis represents the number of samples in the test set.

[0083] By collecting multi-source historical operating data from HVAC monitoring systems and indoor and outdoor sensor networks, covering typical operating conditions such as summer peak, transition season, and winter, training and test datasets are constructed. The input vector consists of 5 types of environmental parameters, and the output vector consists of 2 types of equipment operating parameters. After data cleaning to remove outliers caused by sensor malfunctions, standardized training data is formed, providing high-quality input for the subsequent dual-branch deep and shallow feature fusion neural network model. This ensures that the model can accurately explore the synergistic coupling effect between environmental parameters and cooling load, and achieve accurate load prediction and equipment control optimization.

[0084] Understandably, this embodiment improves the model's environmental adaptability and generalization ability by constructing a training set through the collection of historical data under multiple operating conditions; it enhances the comprehensiveness and accuracy of cold load feature expression through the coordinated collection of five key environmental parameters; it improves dataset quality and model training stability by cleaning and removing outliers; and it improves data scale uniformity and model convergence efficiency through the normalization of chiller pump frequency.

[0085] Based on the above technical solutions, those skilled in the art can make corresponding equivalent improvements according to the specific characteristics of the application scenario or system requirements. For example, the threshold cleaning method can be replaced with an anomaly detection algorithm based on statistical distribution; or linear normalization can be replaced with a minimum-maximum scaling method to adapt to different data distributions; or a sliding window can be used to process time series data to capture dynamic changes.

[0086] Preferably, in step S200, the dual-branch deep and shallow feature fusion neural network model includes a chiller load prediction branch and a chilled pump frequency prediction branch. The steps for jointly predicting the chiller load and chilled pump frequency through the feature fusion mechanism include:

[0087] Step S210: Extract shallow environmental features based on standardized training data, and transform the shallow environmental features into deep environmental features through neural network transformation;

[0088] Step S220: Input the deep environmental characteristics into the chiller load prediction branch, and output the chiller load prediction value and deep chiller characteristics;

[0089] Step S230: Fuse the shallow environment features, deep environment features, and deep chiller features to generate fused features;

[0090] Step S240: Input the fused features into the cryogenic pump model and output the predicted cryogenic pump frequency.

[0091] For details, see Figure 3In a specific embodiment of the present invention, the dual-branch deep and shallow feature fusion neural network model includes a cooling load common feature extraction module, a chiller load prediction branch, a deep and shallow feature fusion module, and a chilled pump frequency prediction branch. The cooling load common feature extraction module receives preprocessed five-dimensional environmental parameter input, including outdoor temperature. T out Outdoor relative humidity RH out Surface radiation intensity of buildings I rad Total number of people indoors N people Total opening of indoor terminals O terminal The system employs a multi-layer fully connected neural network to perform nonlinear transformation and feature abstraction on the original environmental parameters, learning the complex coupling relationship between multiple environmental parameters and cooling load, and outputting high-dimensional deep environmental features. The chiller load prediction branch takes these deep environmental features as input, performs feature compression and correlation mining through a fully connected neural network, and outputs a predicted chiller load value, expressed as a percentage of the chiller's actual cooling capacity to its rated cooling capacity. Simultaneously, it extracts deep chiller features from the network's hidden layers, which contain equipment operating status information. The shallow and deep feature fusion module fuses the original shallow environmental features, deep environmental features, and deep chiller features. The shallow environmental features retain real-time details of the original environmental parameters, the deep environmental features provide high-level abstract correlations, and the deep chiller features provide equipment status. The module achieves dimensionality matching through a projection layer, then uses a residual connection mechanism for superposition and fusion, and finally uses an activation function to generate fused features containing environmental details and equipment status. The chilled water pump frequency prediction branch takes the fused features as input, uses a neural network to capture the coupling relationship between environmental demand and equipment status, and outputs a normalized predicted chilled water pump frequency value. The entire model's data flow starts from the input of environmental parameters, goes through feature extraction, load prediction, feature fusion to frequency prediction, forming a coherent and coordinated control process to ensure the dynamic balance between cooling capacity and flow rate.

[0092] Understandably, this embodiment uses a common feature extraction module for cold load to uniformly process environmental parameters, thereby improving feature consistency and model efficiency; it uses dual-branch collaborative prediction to achieve load matching between the host and the pump, thereby improving the system's global optimization capability; it uses a deep and shallow feature fusion mechanism to integrate environmental details and equipment status, thereby improving the model's adaptability to complex working conditions; and it uses end-to-end neural network modeling to reduce manual intervention and improve the accuracy of intelligent control.

[0093] Based on the above technical solutions, those skilled in the art can make corresponding equivalent improvements according to the specific characteristics of the application scenario or system requirements. For example, the fully connected neural network can be replaced with a graph neural network to handle the topological relationship between parameters; or the residual connection mechanism can be replaced with an attention mechanism to focus on key features; or a convolutional neural network can be used to extract spatial features to enhance local correlation; or a single fusion module can be changed into a multi-level fusion structure to improve feature expression capabilities.

[0094] Preferably, step S210 includes:

[0095] Step S211: Input the preprocessed environmental parameters as shallow environmental features, which include outdoor temperature, outdoor relative humidity, solar radiation intensity, number of people indoors, and indoor terminal opening degree.

[0096] Step S212: The shallow environmental features are transformed nonlinearly through the cold load common feature extraction module. The cold load common feature extraction module adopts a multi-layer fully connected neural network structure, which includes an input layer, at least one hidden layer and an output layer.

[0097] Step S213: Mind the correlation between environmental parameters and cooling load through the activation function of the hidden layer, and output deep environmental features.

[0098] For details, see Figure 4 In a specific embodiment of the present invention, shallow and deep environmental features are extracted from standardized data through a cold load common feature extraction module. The module adopts a fully connected neural network design, takes multi-dimensional environmental parameters as input, and mines the complex correlation between parameters through nonlinear transformation, outputting high-level features with strong representation capabilities. This provides high-quality and robust feature input for the dual-branch prediction network, avoiding prediction deviations caused by insufficient feature quality in subsequent equipment control models.

[0099] First, the preprocessed environmental parameters are used as shallow environmental feature inputs, including outdoor temperature. T out Outdoor relative humidity RH out Surface radiation intensity of buildings I rad Total number of people indoors N people Total opening of indoor terminals O terminalThese parameters reflect environmental heat level, latent heat load, solar heat gain, occupant heat load, and terminal cooling demand, respectively. Data is collected through a sensor network and, after cleaning and standardization, forms a five-dimensional feature vector. Next, a common feature extraction module for cooling load performs a nonlinear transformation on the shallow environmental features. This module employs a multi-layer fully connected neural network structure, including an input layer that receives the five-dimensional shallow features, and at least one hidden layer for feature abstraction and correlation mining. The hidden layer uses an activation function to process the linearly transformed features, capturing the complex coupling relationship between environmental parameters and cooling load, such as the synergistic effect of temperature and humidity, and the superimposed influence of radiation intensity and occupant density. The output layer generates high-dimensional deep environmental features that accurately reflect the dynamic changes in cooling load. Finally, the activation function of the hidden layer achieves nonlinear mapping, and the output deep environmental features serve as the unified input to a dual-branch prediction network, providing a high-quality feature foundation for the subsequent coordinated control of the chiller and chilled water pumps.

[0100] The following example illustrates a possible implementation where the cold load common feature extraction module employs a two-layer fully connected neural network structure. Through a gradual transformation from the input layer to hidden layer 1, hidden layer 2, and output layer, it achieves a non-linear mapping from the original shallow environmental features to deep environmental features. The number of neurons, activation functions, functional localization, and parameter configurations for each layer are shown in Table 1 below.

[0101] Table 1. Structure and parameters of the common feature extraction module for cooling load.

[0102]

[0103] Among them, the common feature extraction model of cooling load is the first i The mathematical expressions for the layer input and output are:

[0104]

[0105] In the formula, Indicates the first i The environmental feature vector input to the layer; Indicates the first i +1 layer input environment feature vector, i.e. the... i The environmental feature vector output by the layer; This represents the activation function. This indicates a batch normalization operation; , This represents a trainable weight matrix; , This represents a trainable bias vector.

[0106] The core training objective of the common feature extraction module for cooling load is to ensure that the output high-level feature vector has a strong linear correlation with the actual cooling load. Furthermore, when input into the bi-branch model, this reduces the prediction errors for equipment load and operating degree compared to directly using the original environmental parameter features. This module effectively extracts the hidden correlation between environmental features and cooling load, providing high-quality feature input for subsequent bi-branch prediction. Secondly, as the "core of feature preprocessing" in the bi-branch prediction network, the common feature extraction module avoids redundant learning, reduces the total number of model parameters, and lowers the computational resource consumption for training and deployment through unified feature extraction. It is highly flexible, and the features output by the module are universal and can be directly transferred to chilled water systems of different buildings and scales.

[0107] Understandably, this embodiment uses a multi-layer fully connected neural network structure to mine the nonlinear correlation of environmental parameters, thereby improving feature abstraction capabilities and representation quality; it avoids redundant calculations and improves model training efficiency and deployment convenience by using a unified feature extraction module; and it provides consistent input for dual-branch prediction by using deep environmental feature output, thereby improving the accuracy of system collaborative control.

[0108] Based on the above technical solutions, those skilled in the art can make corresponding equivalent improvements according to the specific characteristics of the application scenario or system requirements. For example, they can replace the fully connected neural network with a convolutional neural network to enhance the local feature extraction capability; or replace the ReLU activation function with LeakyReLU to retain negative value information.

[0109] Preferably, step S220 includes:

[0110] Step S221: Input the deep environmental features into the neural network of the chiller load prediction branch. The neural network includes an input layer, at least one hidden layer and an output layer.

[0111] Step S222: Through nonlinear transformation of the hidden layer, the deep environmental features are compressed and correlations are mined to generate intermediate features;

[0112] Step S223: The output layer uses an activation function to generate a predicted load value for the chiller unit and outputs the intermediate features as deep chiller unit features.

[0113] For details, see Figure 5 In a specific embodiment of the present invention, the deep environmental features output by the cold load common feature extraction model are... F env Input the chiller load prediction branch to predict the optimal load for the chiller. y chiller Its core function is to accurately match the cooling capacity output of the chiller unit according to changes in environmental factors, so as to provide stable low-temperature chilled water for the chilled water system.

[0114] The following example illustrates that, in one possible embodiment, the neural network for the chiller load prediction branch adopts a three-layer fully connected structure. The design concept of each layer is "feature compression - association mining - load output". The number of neurons, activation functions, functional localization, and parameter configurations of each layer are shown in Table 2 below:

[0115] Table 2 Model structure and parameters of chiller load prediction branch

[0116]

[0117] The mathematical expression for the first hidden layer of the chiller load prediction branch model is as follows:

[0118]

[0119] In the formula, This represents the feature information in the first hidden layer of the chiller load prediction branch layer output; This represents the deep environmental features output by the common feature extraction model for cooling load; This represents the activation function. This indicates a batch normalization operation; , This represents a trainable weight matrix; , This represents a trainable bias vector.

[0120] The mathematical expression for the i-th hidden layer of the chiller load prediction branch model is:

[0121]

[0122] In the formula, This represents the input feature vector of the i-th hidden layer; This represents the feature vector output by the i-th hidden layer, i.e., the... i +1 hidden layer input feature vector.

[0123] First, the input layer receives 64-dimensional deep environmental features. The first hidden layer contains 32 neurons, which undergo linear transformation using a weight matrix and bias vector, followed by batch normalization and ReLU activation to achieve non-linear transformation, compressing the feature dimension to 32 dimensions and generating the first layer of intermediate features. The second hidden layer contains 16 neurons, which further perform correlation mining and dimensionality compression on the features, outputting 16-dimensional intermediate features. These intermediate features contain information on the correlation between the environment and the load, as well as characteristics of equipment operating status. The output layer uses a Sigmoid activation function to map the 16-dimensional intermediate features to predicted chiller load values ​​between 0 and 1. , This represents the percentage ratio of the actual cooling capacity of the chiller to its rated cooling capacity. Simultaneously, the 16-dimensional intermediate features output from the second hidden layer are used as deep chiller feature outputs; these features will be used in subsequent feature fusion modules.

[0124] Understandably, this embodiment improves prediction accuracy and operational adaptability by mining the correlation between environmental features and load through nonlinear transformation of the hidden layer; it improves training efficiency and model convergence speed by stabilizing feature distribution through batch normalization operation; it improves the safety and rationality of control commands by constraining the load output range through the output layer activation function; and it improves the integration of subsequent collaborative control by using intermediate features as equipment status output.

[0125] Preferably, step S230 includes:

[0126] Step S231: Map the shallow environmental features to the same dimension as the deep environmental features through the projection layer to achieve dimension matching;

[0127] Step S232: Using a residual connection mechanism, the mapped shallow environment features, deep environment features, and deep chiller unit features are superimposed and fused.

[0128] Step S233: Apply an activation function to the superimposed and fused features to perform a nonlinear transformation, generating fused features that include environmental details and device status information.

[0129] For details, see Figure 6 In a specific embodiment of the present invention, the shallow and deep feature fusion module achieves efficient integration of environmental features and device status through three core steps. First, dimensional matching processing is performed, mapping the original shallow environmental features, which include five dimensions: outdoor temperature, humidity, solar radiation, indoor personnel density, and terminal opening, to a 64-dimensional feature space with the same dimensions as the deep environmental features through a 1×1 projection layer. The mapping formula is as follows:

[0130]

[0131] In the formula, This represents the feature vector of the shallow environment. This represents the shallow feature vector after mapping; This represents the 64×5 weight matrix of the projection layer, where the number of rows corresponds to the target dimension of 64 and the number of columns corresponds to the input shallow feature dimension of 5. Each element W [i,j] Representing shallow features j After mapping the dimension, the first i Contribution weights of dimensions; This represents the 64-dimensional bias vector of the projection layer, with each element... b proj[i]Used to adjust the mapped first i The baseline value of the dimension is used to compensate for system deviations during the linear transformation process.

[0132] The core of the mapping formula is to distribute the information of 5-dimensional shallow features to 64-dimensional space through a learnable weight matrix. This not only completes dimensional matching, but also allows the weights to automatically focus on key information in the shallow features through training, such as environmental parameters that are strongly correlated with real-time load in HVAC systems, laying the foundation for subsequent fusion of deep and shallow features.

[0133] Next, residual fusion is performed to map the shallow environmental features. with deep features The fusion and overlay are performed using the following formula:

[0134]

[0135] In the formula, Represents the fused feature vector; Represents deep feature vectors, deep features It consists of 64-dimensional deep environmental features and 16-dimensional deep chiller features. The deep chiller features are enhanced to 64 dimensions through linear transformation to ensure dimensional consistency. This represents the shallow feature vector after mapping.

[0136] Using a residual connection mechanism can prevent deep features from overriding shallow details. For example, when the outdoor temperature rises by 5°C, the real-time changes captured by shallow features can be directly transmitted to the chiller load prediction branch without going through multiple nonlinear transformations, which can improve the system's response timeliness.

[0137] Finally, a nonlinear activation function is used to fuse the features. Feature activation is performed using the following formula:

[0138]

[0139] In the formula, This represents the fused feature vector after activation; This represents the inactive fusion feature vector; This represents a trainable bias vector.

[0140] The activation function parameter α was set to 0.1 to preserve negative feature information and enhance the model's expressive power. The final generated 64-dimensional fused features... It contains real-time environmental details and in-depth equipment status information to provide a basis for predicting the frequency of chilled pumps. The entire process ensures feature compatibility through dimensional matching, ensures information integrity through residual connections, and enhances nonlinear expressive power through activation functions, enabling the chilled pump model to both quickly respond to environmental changes and deeply understand the collaborative patterns of equipment.

[0141] Understandably, this embodiment achieves feature space alignment through projection layer dimension matching, improving the compatibility and efficiency of the fusion process; retains shallow environmental details through the residual connection mechanism, improving the model's response speed to sudden operating conditions; introduces nonlinear transformation through the LeakyReLU activation function, improving feature representation capabilities and model adaptability; and improves the accuracy of the refrigeration pump branch's understanding of the environment and equipment status through multi-source feature superposition and fusion.

[0142] Based on the above technical solutions, those skilled in the art can make corresponding equivalent improvements according to the specific characteristics of the application scenario or system requirements. For example, feature splicing can be used instead of superposition and fusion to maintain the independence of each feature stream; or the 1×1 projection layer can be replaced with a fully connected layer to enhance the feature transformation capability.

[0143] Preferably, step S240 includes:

[0144] Step S241: Input the fused features into the neural network of the cryopump frequency prediction branch, the neural network including an input layer, at least one hidden layer and an output layer;

[0145] Step S242: The fusion features are processed by nonlinear transformation of the hidden layer to capture the coupling relationship between environmental requirements and equipment status;

[0146] Step S243: The output layer uses an activation function to generate normalized predicted values ​​for the refrigeration pump frequency.

[0147] For details, see Figure 7 In a specific embodiment of the present invention, the deep and shallow feature fusion module will activate the fused feature vector. The neural network output to the chilled water pump frequency prediction branch adopts a three-layer fully connected neural network structure, with a higher number of neurons than the chiller unit branch, to adapt to the complex coupling relationship of "environment-unit-pump". The number of neurons, activation functions, functional positioning and parameter configuration of each layer are shown in Table 3 below:

[0148] Table 3. Model structure and parameters of the frequency prediction branch for chilled pumps

[0149]

[0150] The mathematical expression for the first hidden layer of the chilled pump frequency prediction branch model is:

[0151]

[0152] In the formula, This represents the feature information in the first hidden layer of the output of the refrigeration pump frequency prediction branch model; This represents the fused feature vector of the input activation; This represents the activation function. This indicates a batch normalization operation; , This represents a trainable weight matrix; , This represents a trainable bias vector.

[0153] The mathematical expression for the i-th hidden layer of the chilled pump frequency prediction branch model is:

[0154]

[0155] In the formula, This represents the input feature vector of the i-th hidden layer; This represents the feature vector output by the i-th hidden layer, i.e., the... i +1 hidden layer input feature vector.

[0156] In this embodiment, the chilled pump frequency prediction branch receives 64-dimensional fused features from the deep and shallow feature fusion module. The first hidden layer contains 64 neurons, which fuse the input features through a weight matrix and a bias vector. A linear transformation is performed, followed by batch normalization and the introduction of nonlinearity using the ReLU activation function to initially capture the coupling relationship between environmental requirements and equipment status. The second hidden layer contains 32 neurons, further compressing the feature dimensions and deepening feature abstraction. Nonlinear transformation is used to mine the flow regulation patterns hidden in the fused features, such as identifying the pump frequency requirements corresponding to high loads of the unit under high temperature and humidity conditions. The output layer maps the 32-dimensional features to normalized refrigeration pump frequency prediction values ​​using the Sigmoid activation function. , The value ranges from 0 to 1, representing the normalized load rate of the refrigeration pump, which can be converted into an actual frequency control command through inverse normalization.

[0157] Understandably, this embodiment improves the accuracy of frequency prediction and adaptability to operating conditions by integrating environmental and equipment status information through fusion feature inputs; it enhances the model's responsiveness to dynamic demands by capturing complex coupling relationships through nonlinear transformation of the hidden layer; and it improves the safety and reliability of control commands by generating normalized values ​​through the Sigmoid activation function of the output layer.

[0158] Based on the above technical solutions, those skilled in the art can make corresponding equivalent improvements according to the specific characteristics of the application scenario or system requirements. For example, they can replace the fully connected neural network with a convolutional neural network to enhance the local feature extraction capability; or replace the ReLU activation function with LeakyReLU to improve the processing of negative value information.

[0159] Preferably, step S400 includes:

[0160] Step S410: Based on the deviation between the supply and return water temperature difference and the ideal temperature difference, proportionally correct the predicted value of the chilled water pump frequency;

[0161] Step S420: Perform boundary constraint processing on the corrected chiller load prediction value and chilled pump frequency prediction value to ensure that they are within the safe range of equipment operation;

[0162] Step S430: Output the final control command that satisfies the physical constraints.

[0163] For details, see Figure 8 In one specific embodiment of the present invention, a constraint correction module is also provided to optimize the model predictions based on physical constraints, ensuring that the output conforms to both data patterns and physical laws. One of the physical parameters directly related to system safety and energy consumption is the current temperature difference between the chilled water supply and return water. Δt 当前 When the temperature difference is less than the ideal temperature difference Δt 理想 When the temperature difference is greater than the ideal temperature difference, it indicates flow redundancy and the pump frequency needs to be reduced. Δt 理想 This indicates insufficient flow, requiring an increase in pump frequency. Simultaneously, to avoid frequent equipment start-ups and overloads, the load output of the chiller unit model and the output frequency of the chilled water pump model need to be limited to their rated range. The specific steps are shown in the following formula:

[0164]

[0165] In the formula, This represents the corrected predicted frequency of the refrigeration pump; This represents the normalized frequency value of the original output of the refrigeration pump frequency prediction branch model; It is the coefficient for proportional adjustment. This represents the correction value for the output frequency, used to dynamically adjust the frequency value based on the temperature difference deviation. This represents the output frequency at the current moment (time step i); This represents the output frequency at the previous moment (time step i-1); This indicates the currently measured temperature difference between the chilled water supply and return water. This indicates the ideal or set temperature difference between the chilled water supply and return water.

[0166] Then, the output load of the chiller load prediction branch model is analyzed. y chiller Perform boundary screening; if the value exceeds the range of the device output value, correct it directly.

[0167]

[0168] In the formula, This represents the corrected chiller unit load prediction value; This represents the load prediction value of the original output of the chiller load prediction branch model; This indicates the maximum allowable load on the chiller unit; This indicates the minimum allowable load on the chiller unit.

[0169] Meanwhile, the output frequency of the refrigeration pump branch model after temperature difference correction was also adjusted. Perform boundary screening; if it exceeds the physical range, correct it directly.

[0170]

[0171] In the formula, This represents the final corrected refrigeration pump frequency value; This represents the normalized upper bound boundary value for the frequency of the refrigeration pump, typically 1; This represents the normalized lower limit boundary value of the refrigeration pump frequency, which is usually 0.

[0172] Understandably, this embodiment dynamically adjusts the flow rate by correcting the temperature difference deviation ratio, thereby improving system energy efficiency and response accuracy; and limits the operating range of the equipment by using boundary constraint processing, thereby improving safety and stability.

[0173] Based on the above technical solutions, those skilled in the art can make corresponding equivalent improvements according to the specific characteristics of the application scenario or system requirements. For example, the fixed boundary can be changed to an adaptive range adjustment to adapt to equipment aging; or predictive control can be introduced to optimize and correct parameters to improve long-term performance; or multi-objective optimization can be adopted to balance energy consumption and comfort to expand application scenarios.

[0174] Preferably, in step S200, the step of training the dual-branch deep and shallow feature fusion neural network model based on standardized training data until the model converges includes:

[0175] Step S250: Based on standardized training data, perform the forward propagation process of the dual-branch deep and shallow feature fusion network model to obtain the joint predicted value of chiller load and chilled pump frequency, and construct a comprehensive loss function, which includes chiller load prediction loss, chilled pump frequency prediction loss and system energy consumption loss.

[0176] Step S260: Calculate the gradient of the comprehensive loss function with respect to the model parameters using the gradient descent algorithm, and optimize the parameters through backpropagation to minimize the comprehensive loss function;

[0177] Step S270: Iteratively execute the forward and backward propagation processes. When the loss of the comprehensive loss function satisfies the preset convergence condition, stop training and save the optimal model parameters.

[0178] For details, see Figure 9 In a specific embodiment of the present invention, the training process of the dual-branch deep and shallow feature fusion neural network model revolves around minimizing the comprehensive loss function, aiming to balance prediction accuracy and system energy efficiency. First, the comprehensive loss function is constructed, encompassing three main aspects: feature extraction quality, prediction accuracy of the chiller and chilled pumps, and system energy consumption. The overall optimization objective expression is shown below:

[0179]

[0180] In the formula, Indicates the parameters of the common feature extraction module. This indicates the parameters of the chilled pump network module. This indicates the parameters of the chiller unit module; A matrix representing the input data. A label matrix representing the actual chiller load and chilled water pump frequency; This indicates the predicted load loss of the chiller unit; This indicates the predicted loss due to the frequency of the chilled pump. Indicates system energy loss; The training optimization weight coefficients represent the losses of the chiller unit; The training optimization weights represent the losses of the cryogenic pump; This represents the training optimization weight coefficient for the energy loss term.

[0181] For chiller load prediction loss Mean squared error loss is used:

[0182]

[0183] In the formula, Indicates the first i Predicted load values ​​for chiller units in a sample. Indicates the first i The actual load value of the chiller unit for each sample. This represents the total number of training samples.

[0184] For the predicted loss of chilled pump frequency Similarly, mean squared error loss is used:

[0185]

[0186] In the formula, Indicates the first i Predicted frequency values ​​of the cryopumps for each sample. Indicates the first i The true value of the cryopump frequency for each sample. This represents the total number of training samples.

[0187] For energy consumption optimization losses in chilled water systems, the loss value is... The energy consumption is proportional to the system energy consumption for each sample:

[0188]

[0189]

[0190]

[0191] In the formula, This represents the input feature vector of the i-th sample; express i Based on the input vector under sample working conditions Calculated chiller energy consumption, express i Based on the input vector under sample working conditions Calculated energy consumption of the refrigeration pump; The coefficient of performance (COP) of a chiller is positively correlated with the rated cooling capacity of the equipment. The coefficient of performance (COP) of the chiller and chilled water pump is positively correlated with the rated cooling capacity of the equipment. Indicates based on input feature vector Chiller load prediction branch prediction value; Indicates based on input feature vector The predicted value of the branch of the refrigeration pump frequency prediction; This indicates that the chiller is in the corresponding i Sample working conditions COP The value needs to be obtained by referring to the partial load performance curve provided by the equipment manufacturer; Indicates the refrigeration pump is in the corresponding i Sample working conditions COP For specific values, please refer to the partial load performance curves provided by the equipment manufacturer.

[0192] The parameters of the common feature extraction module affect the subsequent model output through the extracted common features. Therefore, the gradient update method for the parameters of the common feature extraction module based on the overall optimization objective is as follows:

[0193]

[0194] In the formula, This represents the set of trainable parameters for the common feature extraction module of the cold load; Represents the total loss function For parameters The gradient; Represents the loss function For parameters The gradient; Represents the loss function For parameters The gradient; Represents the loss function For parameters The gradient; Indicates loss item Weighting coefficients; Indicates loss item Weighting coefficients; Indicates loss item The weighting coefficients.

[0195] The parameters of the chiller module directly affect the chiller load, the prediction accuracy of the chilled pump, and the energy consumption calculation results. Therefore, the gradient update method for the chiller module parameters is as follows:

[0196]

[0197] In the formula, This represents the set of trainable parameters for the chiller load prediction branch module; Represents the total loss function For parameters The gradient; Represents the loss function For parameters The gradient; Represents the loss function For parameters The gradient; Represents the loss function For parameters The gradient; Indicates loss item Weighting coefficients; Indicates loss item Weighting coefficients; Indicates loss item The weighting coefficients.

[0198] The parameters of the chilled pump module directly affect the prediction accuracy and energy consumption calculation results of the chilled pump. Therefore, the gradient update method for the parameters of the chilled pump network module is as follows:

[0199] .

[0200] In the formula, This represents the set of trainable parameters for the common feature extraction module; Represents the total loss function For parameters The gradient; Represents the loss function For parameters The gradient; Represents the loss function For parameters The gradient; Indicates loss item Weighting coefficients; Indicates loss item The weighting coefficients.

[0201] The model training process uses the Adam optimizer with an initial learning rate of 0.001, which is then dynamically adjusted. The learning rate decays by 50% after 10 epochs of validation loss stagnation. An early stopping strategy is implemented, terminating training after 20 epochs of validation loss stagnation to restore optimal weights and prevent overfitting. The training epochs are set to 150, with a batch size of 32. The training set is divided into training and validation subsets in an 8:2 ratio. The entire training process utilizes a multi-objective loss function to ensure the model balances prediction accuracy and energy efficiency, an adaptive optimization algorithm to improve training efficiency, and an early stopping mechanism to prevent overfitting, ultimately resulting in a highly generalizable intelligent control model.

[0202] Understandably, this embodiment improves the model's practicality by balancing prediction accuracy and system energy efficiency through a comprehensive loss function; it enhances the model's collaborative performance by optimizing all network parameters through a gradient descent algorithm; it improves the model's generalization ability and training efficiency through validation monitoring and early stopping mechanisms; it enhances adaptability to different application scenarios through multi-objective weighted design; and it improves training stability and convergence speed through batch training and gradient pruning.

[0203] Based on the above technical solutions, those skilled in the art can make corresponding equivalent improvements according to the specific characteristics of the application scenario or system requirements. For example, the mean squared error loss can be replaced with smoothed L1 loss to enhance the robustness of outliers; or the Adam optimizer can be replaced with SGD with momentum to improve convergence accuracy; or a learning rate decay strategy can be used to replace the fixed learning rate to improve the stability in the later stages of training; or a regularization term can be introduced into the loss function to control the model complexity; or cross-validation can be used to replace simple validation set monitoring to optimize early stopping decisions.

[0204] See Figure 10 and Figure 11 The second aspect of this invention proposes a HVAC chilled water collaborative control system based on a dual-branch deep and shallow feature fusion neural network, comprising:

[0205] The data acquisition and preprocessing module is used to acquire historical operating data of the HVAC chilled water system, preprocess the historical operating data to form standardized training data; the historical operating data includes environmental parameters and equipment operating parameters.

[0206] The dual-branch model training module is used to train the dual-branch deep and shallow feature fusion neural network model based on standardized training data until the model converges. The dual-branch deep and shallow feature fusion neural network model includes a chiller load prediction branch and a chilled pump frequency prediction branch, which realizes the joint prediction of chiller load and chilled pump frequency opening through feature fusion mechanism.

[0207] The real-time predictive control module is used to collect environmental parameters of the HVAC chilled water system in real time, preprocess them, and then input them into a trained dual-branch deep and shallow feature fusion neural network model to output the predicted values ​​of chiller load and chilled water pump frequency.

[0208] The constraint correction module is used to correct the predicted load values ​​of the chiller and the predicted frequency values ​​of the chilled water unit based on a preset constraint correction algorithm, and output the corrected control commands.

[0209] A third aspect of the present invention also provides a storage medium storing a HVAC chilled water collaborative control processing program based on a dual-branch deep and shallow feature fusion neural network. When the HVAC chilled water collaborative control program based on the dual-branch deep and shallow feature fusion neural network is executed by a processor, it implements the steps of the HVAC chilled water collaborative control method based on the dual-branch deep and shallow feature fusion neural network as described in any of the above embodiments.

[0210] A fourth aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, it implements the HVAC chilled water collaborative control method based on a dual-branch deep and shallow feature fusion neural network as described in any embodiment of the first aspect.

[0211] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. This processor is the control center of the HVAC chilled water collaborative control system based on a dual-branch deep and shallow feature fusion neural network, connecting various parts of the operating device through various interfaces and lines.

[0212] The memory can be used to store the computer programs and / or modules. The processor, by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory, realizes various functions of the HVAC chilled water collaborative control system based on a dual-branch deep and shallow feature fusion neural network. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0213] Compared with the prior art, the beneficial effects of the present invention include at least the following:

[0214] This invention provides a collaborative control method and system for HVAC chilled water based on a dual-branch shallow and deep feature fusion neural network. The method achieves collaborative prediction between the chiller and chilled water pump through the dual-branch shallow and deep feature fusion neural network, improving the system's global optimization capability and control accuracy. It integrates shallow environmental features with deep equipment states through a feature fusion mechanism, enhancing the model's adaptability to complex operating conditions and predictive timeliness. A physical constraint correction module ensures the output conforms to equipment operating patterns, improving system safety and reliability. Neural network modeling reduces reliance on expert experience, improving intelligent control efficiency and practicality. Training based on historical data enables the model to adapt to changing operating conditions, improving energy-saving effects and generalization ability.

[0215] Furthermore, this invention improves the completeness of the training dataset by collecting multi-dimensional environmental and equipment parameters; improves the quality of the training dataset by cleaning abnormal data; enhances the model's feature representation ability by extracting shallow and deep environmental features; improves the control coordination between devices through dual-branch collaborative prediction; improves the accuracy of the chilled pump model's understanding of complex operating conditions through a deep and shallow feature fusion mechanism; improves the response speed to sudden operating conditions by retaining the original environmental parameters as shallow environmental features; improves the accuracy of cooling load prediction by mining deep correlations between parameters through a common feature extraction module; improves the accuracy of chiller unit load prediction by inputting deep environmental features into the load prediction branch; and improves the accuracy of chiller unit load prediction by outputting intermediate features as... The deep chiller host features are used to improve the information integrity of subsequent feature fusion; dimension matching is achieved through a projection layer to improve the compatibility of feature fusion; the integrity of feature information is improved through a residual connection mechanism; the representational ability of fused features is improved through nonlinear transformation of activation functions; the environmental adaptability of chiller pump control is improved by using the fused feature input chiller pump frequency prediction branch; control accuracy is improved by temperature difference deviation ratio correction; operational safety is improved by boundary constraint processing; system reliability is improved by physical constraint output; multi-objective optimization capability is improved by constructing a comprehensive loss function; training efficiency is improved by optimizing parameters through gradient descent algorithm; and model generalization ability is improved by stopping training through verification loss convergence conditions.

[0216] In summary, the present invention proposes a collaborative control method and system for HVAC chilled water based on a dual-branch deep and shallow feature fusion neural network, which solves the technical problems of insufficient equipment collaboration, inadequate feature extraction, and poor control reliability caused by lack of physical constraints in the prior art.

[0217] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention. Equivalent structural transformations made using the description and drawings of the present invention, or direct / indirect applications in other related technical fields, are all included within the scope of patent protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

[0218] It should be noted that the embodiments implemented in the HVAC chilled water collaborative control system based on the dual-branch deep and shallow feature fusion neural network in this invention can be referenced with the embodiments implemented in the HVAC chilled water collaborative control method based on the dual-branch deep and shallow feature fusion neural network. These embodiments will not be described in detail in this invention.

[0219] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A collaborative control method for HVAC chilled water based on a dual-branch deep and shallow feature fusion neural network, characterized in that, include: Step S100: Obtain historical operating data of the HVAC chilled water system, preprocess the historical operating data to form standardized training data; The historical operational data includes environmental parameters and equipment operating parameters; Step S200: Train the dual-branch deep and shallow feature fusion neural network model based on the standardized training data until the model converges; the dual-branch deep and shallow feature fusion neural network model includes a chiller load prediction branch and a chilled pump frequency prediction branch, and realizes the joint prediction of chiller load and chilled pump frequency opening through feature fusion mechanism. Step S300: Collect environmental parameters of the HVAC chilled water system in real time, preprocess them, and input them into the trained dual-branch deep and shallow feature fusion neural network model to output the chiller load prediction value and the chilled pump frequency prediction value. Step S400: Based on the preset constraint correction algorithm, correct the predicted load value of the chiller and the predicted frequency value of the chilled water pump, and output the corrected control command.

2. The HVAC chilled water collaborative control method based on a dual-branch deep and shallow feature fusion neural network as described in claim 1, characterized in that, Step S100 includes: Step S110: Collect historical operating data of the HVAC chilled water system. The environmental parameters include outdoor temperature, outdoor relative humidity, solar radiation intensity, number of people indoors, and indoor terminal opening. The equipment operating parameters include chiller load rate and chilled water pump frequency. Step S120: Denoise the collected historical operating data to remove outliers caused by sensor malfunctions and invalid data that are outside the physical reasonable range; Step S130: Standardize the cleaned historical running data to form training dataset and test dataset.

3. The HVAC chilled water collaborative control method based on a dual-branch deep and shallow feature fusion neural network as described in claim 2, characterized in that, In step S200, the dual-branch deep and shallow feature fusion neural network model includes a chiller load prediction branch and a chilled pump frequency prediction branch. The steps for jointly predicting the chiller load and chilled pump frequency through the feature fusion mechanism include: Step S210: Extract shallow environmental features based on the standardized training data, and transform the shallow environmental features into deep environmental features through a neural network. Step S220: Input the deep environmental features into the chiller load prediction branch, and output the chiller load prediction value and the deep chiller features; Step S230: The shallow environment features, the deep environment features, and the deep chiller features are fused to generate fused features; Step S240: Input the fused features into the refrigeration pump model and output the predicted refrigeration pump frequency value.

4. The HVAC chilled water collaborative control method based on a dual-branch deep and shallow feature fusion neural network as described in claim 3, characterized in that, Step S210 includes: Step S211: Input the preprocessed environmental parameters as shallow environmental features, which include outdoor temperature, outdoor relative humidity, solar radiation intensity, number of people indoors, and indoor terminal opening degree. Step S212: The shallow environmental features are nonlinearly transformed by the cold load common feature extraction module. The cold load common feature extraction module adopts a multi-layer fully connected neural network structure, which includes an input layer, at least one hidden layer and an output layer. Step S213: Mine the correlation between the environmental parameters and the cooling load through the activation function of the hidden layer, and output the deep environmental features.

5. The HVAC chilled water collaborative control method based on a dual-branch deep and shallow feature fusion neural network as described in claim 3, characterized in that, Step S220 includes: Step S221: Input the deep environmental features into the neural network of the chiller load prediction branch, wherein the neural network includes an input layer, at least one hidden layer and an output layer; Step S222: Through the nonlinear transformation of the hidden layer, the deep environmental features are compressed and correlated to generate intermediate features; Step S223: The output layer uses an activation function to generate a predicted load value for the chiller unit and outputs the intermediate features as deep chiller unit features.

6. The HVAC chilled water collaborative control method based on a dual-branch deep and shallow feature fusion neural network as described in claim 3, characterized in that, Step S230 includes: Step S231: Map the shallow environmental features to the same dimension as the deep environmental features through a projection layer to achieve dimension matching; Step S232: Using a residual connection mechanism, the mapped shallow environmental features, deep environmental features, and deep chiller features are superimposed and fused. Step S233: Apply an activation function to the superimposed and fused features to perform a nonlinear transformation, generating fused features that include environmental details and device status information.

7. The HVAC chilled water collaborative control method based on a dual-branch deep and shallow feature fusion neural network as described in claim 3, characterized in that, Step S240 includes: Step S241: Input the fused features into the neural network of the cryogenic pump frequency prediction branch, the neural network including an input layer, at least one hidden layer and an output layer; Step S242: Process the fused features through the nonlinear transformation of the hidden layer to capture the coupling relationship between environmental requirements and device status; Step S243: The output layer uses an activation function to generate normalized predicted values ​​for the refrigeration pump frequency.

8. The HVAC chilled water collaborative control method based on a dual-branch deep and shallow feature fusion neural network as described in any one of claims 1 to 7, characterized in that, Step S400 includes: Step S410: Based on the deviation between the supply and return water temperature difference and the ideal temperature difference, the predicted value of the chilled water pump frequency is proportionally corrected; Step S420: Perform boundary constraint processing on the corrected chiller load prediction value and chilled pump frequency prediction value to ensure that they are within the safe range of equipment operation; Step S430: Output the final control command that satisfies the physical constraints.

9. The HVAC chilled water collaborative control method based on a dual-branch deep and shallow feature fusion neural network as described in any one of claims 1 to 7, characterized in that, In step S200, the step of training the dual-branch deep and shallow feature fusion neural network model based on the standardized training data until the model converges includes: Step S250: Based on the standardized training data, perform the forward propagation process of the dual-branch deep and shallow feature fusion neural network model to obtain the joint predicted value of chiller load and chilled pump frequency, and construct a comprehensive loss function, which includes chiller load prediction loss, chilled pump frequency prediction loss and system energy consumption loss. Step S260: Calculate the gradient of the comprehensive loss function with respect to the model parameters using the gradient descent algorithm, and optimize the parameters through backpropagation to minimize the comprehensive loss function; Step S270: Iteratively execute the forward propagation and backward propagation process. When the loss of the integrated loss function is verified to meet the preset convergence condition, stop training and save the optimal model parameters.

10. A HVAC chilled water collaborative control system based on a dual-branch deep and shallow feature fusion neural network, characterized in that, include: The data acquisition and preprocessing module is used to acquire historical operating data of the HVAC chilled water system, preprocess the historical operating data, and form standardized training data. The historical operational data includes environmental parameters and equipment operating parameters; The dual-branch model training module is used to train the dual-branch deep and shallow feature fusion neural network model based on the standardized training data until the model converges; the dual-branch deep and shallow feature fusion neural network model includes a chiller load prediction branch and a chilled pump frequency prediction branch, and realizes the joint prediction of the chiller load and chilled pump frequency opening through a feature fusion mechanism. The real-time predictive control module is used to collect environmental parameters of the HVAC chilled water system in real time, preprocess them, and then input them into a trained dual-branch deep and shallow feature fusion neural network model to output the predicted values ​​of chiller load and chilled water pump frequency. The constraint correction module is used to correct the predicted load value of the chiller and the predicted frequency value of the chilled water unit based on a preset constraint correction algorithm, and output the corrected control command.

Images