Intelligent water quality monitoring and operation regulation and control method and system for central air conditioning system

Abstract

The invention discloses an intelligent water quality monitoring and operation regulation and control method and system for a central air-conditioning system. The method comprises the following steps: collecting multi-dimensional data in real time; the method comprises the following steps: extracting short-term local features through a multi-layer LSTM gating mechanism, capturing long-term dependence by using a multi-head self-attention mechanism, establishing a hybrid network model, optimizing hyper-parameters of the hybrid network model by using an improved IWOA whale optimization algorithm, predicting multi-dimensional data through a target hybrid network model, and constructing a GNN graph neural network. Abstracting a water system of the central air conditioner into a topological graph of nodes and edges, and modeling a dynamic coupling relationship between equipment; and initially embedding the spatio-temporal feature vectors as nodes of a GNN graph neural network, aggregating neighbor information by using each node, updating node states through a GCN graph convolutional layer, and outputting an operation regulation and control scheme. And the control efficiency is improved while the resource utilization rate is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of water quality processing of central air conditioning systems, and in particular to a central air conditioning system water quality intelligent monitoring and operation regulation method and system. BACKGROUND

[0002] Central air conditioning systems are prone to heat exchange efficiency decline, equipment life shortening and energy consumption increase due to problems such as circulating water fouling, microbial breeding and corrosion in long-term operation. The existing technology mostly uses single parameter monitoring or fixed rule regulation, which has the problems of incomplete monitoring: not integrating multi-dimensional water quality parameters (turbidity, microbial content, dissolved oxygen, etc.) and operation state (flow, temperature), and being difficult to accurately evaluate water quality risk. And usually relying on manual intervention or fixed period operation, unable to respond to water quality changes in real time, resulting in low processing efficiency. SUMMARY

[0003] The purpose of the present application is to solve the above problems, and a central air conditioning system water quality intelligent monitoring and operation regulation method is designed.

[0004] To achieve the above purpose, the technical scheme of the present application is that, further, in the central air conditioning system water quality intelligent monitoring and operation regulation method, the central air conditioning system water quality intelligent monitoring and operation regulation method comprises the following steps: Real-time collection of water quality data, environmental parameters and operation state data of central air conditioning equipment, preprocessing of the collected data to obtain multi-dimensional data; Extracting short-term local features through the gating mechanism of multi-layer LSTM, capturing long-term dependencies using the multi-head self-attention mechanism of Transformer, establishing an LSTM+Transformer hybrid network model, Optimizing the hyperparameters of the hybrid network model using the improved IWOA whale optimization algorithm to obtain a target hybrid network model, predicting the multi-dimensional data through the target hybrid network model, and outputting a spatio-temporal feature vector; Building a GNN graph neural network, abstracting the water system of the central air conditioning as a topological graph of nodes and edges, and modeling the dynamic coupling relationship between devices; The spatio-temporal feature vector is used as the initial embedding of the nodes of the GNN graph neural network, the neighbor information of each node is aggregated, the node state is updated through the GCN graph convolution layer, and an operation regulation scheme is output.

[0005] Further, in the central air conditioning system water quality intelligent monitoring and operation regulation method, the real-time collection of water quality data, environmental parameters and operation state data of central air conditioning equipment, and the preprocessing of the collected data to obtain multi-dimensional data, comprises: Real-time data collection of water quality, environmental parameters, and operating status data of central air conditioning equipment; linear interpolation is used to fill in missing data to obtain cleaning data. Outliers in the cleaned data are detected using the box plot method. The quartiles of the data are calculated, and data that exceeds the range are identified as outliers and deleted to obtain the processed data. After normalizing the processed data, the water quality data, environmental parameters, and air conditioning operation status data are aligned according to the time series to obtain multi-dimensional data indexed by time.

[0006] Furthermore, in the aforementioned method for intelligent water quality monitoring and operation control of a central air conditioning system, the step of extracting short-term local features through a multi-layer LSTM gating mechanism and capturing long-term dependencies using a Transformer multi-head self-attention mechanism to establish an LSTM+Transformer hybrid network model includes: Input sensor time-series data and extract short-term local features through the forget gate, input gate and output gate of a multi-layer LSTM; By using a forget gate to filter out irrelevant noise, the output gate generates a hidden state sequence and preserves the short-term fluctuation patterns of water quality parameters.

[0007] Furthermore, in the aforementioned method for intelligent monitoring and operation control of water quality in a central air conditioning system, the step of extracting short-term local features through the gating mechanism of a multi-layer LSTM and capturing long-term dependencies using the multi-head self-attention mechanism of a Transformer to establish an LSTM+Transformer hybrid network model further includes: Input the hidden state sequence output by LSTM into Transformer, and capture long-term dependencies through multi-head self-attention mechanism; The correlation between different time steps is quantified by self-attention weights, temporal sequence information is injected by position encoding, and local features of LSTM and global features of Transformer are weighted and fused to generate spatiotemporal feature vectors.

[0008] Furthermore, in the aforementioned method for intelligent water quality monitoring and operation control of a central air conditioning system, the improved IWOA whale optimization algorithm is used to optimize the hyperparameters of the hybrid network model to obtain a target hybrid network model. This target hybrid network model is then used to predict the multi-dimensional data and output a spatiotemporal feature vector, including: A certain number of hyperparameter combinations are randomly generated as the initial population, with each individual representing a set of hyperparameters; each set of hyperparameters is substituted into the LSTM+Transformer hybrid network model, and the prediction error of the model on the training set is used as the fitness value. The population is updated using a search strategy based on the improved IWOA whale optimization algorithm, including predation, encirclement, and search operations, to find a better combination of hyperparameters. The optimal combination of hyperparameters obtained when the number of iterations reaches the preset maximum value is the target hyperparameter. Substituting the target hyperparameter into the model yields the target hybrid network model.

[0009] Furthermore, in the aforementioned method for intelligent monitoring and operation control of water quality in a central air conditioning system, the construction of a GNN graph neural network, which abstracts the water system of the central air conditioning system as a topological graph of nodes and edges, and models the dynamic coupling relationships between devices, includes: The initial attribute features of the receiving nodes, including the encoded values ​​of basic attributes and the standardized values ​​of real-time state attributes, are transformed into a 128-dimensional initial feature vector for each node. Two graph convolutional layers are set up to aggregate node information. The first layer focuses on the information interaction between directly adjacent nodes, and the second layer expands the aggregation scope to include the information of indirectly connected nodes. A non-linear activation function is introduced after each graph convolution to enhance the model's expressive power and to process the aggregated features.

[0010] Furthermore, in the aforementioned method for intelligent water quality monitoring and operation control of a central air conditioning system, the step of using the spatiotemporal feature vector as the initial embedding of nodes in the GNN graph neural network, aggregating neighbor information for each node, updating node states through a GCN graph convolutional layer, and outputting an operation control scheme includes: Based on the final node state, the GNN (Graph Neural Network) outputs a device control scheme, which includes at least pump speed adjustment, valve opening adjustment, heat exchanger area adjustment, and filter replacement.

[0011] Furthermore, in a central air conditioning system water quality intelligent monitoring and operation control device, the central air conditioning system water quality intelligent monitoring and operation control device includes the following modules: The multi-dimensional data acquisition module is used to collect water quality data, environmental parameters and operating status data of central air conditioning equipment in real time, and to preprocess the collected data to obtain multi-dimensional data. The hybrid model building module is used to extract short-term local features through the gating mechanism of a multi-layer LSTM and capture long-term dependencies using the multi-head self-attention mechanism of the Transformer, thus building an LSTM+Transformer hybrid network model. The feature vector generation module is used to optimize the hyperparameters of the hybrid network model using the improved IWOA whale optimization algorithm to obtain the target hybrid network model. The target hybrid network model is then used to predict the multi-dimensional data and output spatiotemporal feature vectors. The graph neural network building module is used to construct a GNN graph neural network, which abstracts the water system of the central air conditioning system into a topological graph of nodes and edges, and models the dynamic coupling relationship between equipment. The regulation scheme generation module is used to use the spatiotemporal feature vector as the initial embedding of the nodes of the GNN graph neural network, aggregate the neighbor information of each node, update the node state through the GCN graph convolutional layer, and output the operation regulation scheme.

[0012] Furthermore, in the equipment for implementing the above-mentioned intelligent water quality monitoring and operation control method for a central air conditioning system, the graph neural network establishment module includes the following sub-modules: The receiving submodule is used to receive the initial attribute features of the nodes, including the encoded values ​​of the basic attributes and the standardized values ​​of the real-time state attributes. Each node is transformed into a 128-dimensional initial feature vector. The configuration submodule is used to set up two graph convolutional layers to aggregate node information. The first layer focuses on the information interaction between directly adjacent nodes, and the second layer expands the aggregation scope to include information of indirectly connected nodes. A submodule is introduced to introduce a non-linear activation function after each graph convolution to enhance the model's expressive power and to process the aggregated features.

[0013] Furthermore, in the equipment for implementing the above-mentioned intelligent water quality monitoring and operation control method for a central air conditioning system, the control scheme generation module includes the following sub-modules: The output submodule is used by the GNN graph neural network to output equipment control schemes based on the final node state, including at least pump speed adjustment, valve opening adjustment, heat exchanger area adjustment, and filter replacement.

[0014] Its beneficial effects are as follows: 1. It can comprehensively consider the mutual influence between equipment, resulting in more scientific and coordinated control schemes. For example, when adjusting the water pump speed, it comprehensively considers the heat exchanger efficiency, filter clogging, etc., avoiding increased energy consumption caused by frequent adjustments of a single device. This improves resource utilization while enhancing control efficiency. 2. It effectively reduces equipment failures caused by water quality issues. Simultaneously, by optimizing equipment operating parameters, it ensures equipment operates under reasonable conditions, reducing wear and aging, extending equipment lifespan, extending replacement cycles, and lowering replacement costs and downtime. 3. The system can automatically generate equipment operation reports and control records, providing managers with comprehensive and accurate equipment operation information, facilitating data analysis and decision-making, and improving the level of management precision and work efficiency. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of a method for intelligent water quality monitoring and operation control of a central air conditioning system according to the present invention; Figure 2 This is a schematic diagram of a second embodiment of a method for intelligent water quality monitoring and operation control of a central air conditioning system according to the present invention; Figure 3 This is a schematic diagram of the first embodiment of a central air conditioning system water quality intelligent monitoring and operation control device according to the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "one," "an," "say," and "this" used herein may also include plural forms. It should be further understood that the terminology used in this specification includes the presence of the stated feature, integer, step, operation, element, and / or component, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings, such as... Figure 1 As shown, a method for intelligent monitoring and operation control of water quality in a central air conditioning system includes the following steps: Step 101: Collect water quality data, environmental parameters and operating status data of the central air conditioning equipment in real time, and preprocess the collected data to obtain multi-dimensional data; Specifically, in this embodiment, water quality data, environmental parameters, and operating status data of the central air conditioning equipment are collected in real time, and linear interpolation is used to fill in the missing data to obtain cleaning data; Outliers in the cleaned data are detected using the box plot method. The quartiles of the data are calculated, and data that exceeds the range are identified as outliers and deleted to obtain the processed data. After normalizing the processed data, the water quality data, environmental parameters, and air conditioning operation status data are aligned according to the time series to obtain multi-dimensional data indexed by time.

[0020] Specifically, (a) Data collection content and methods Water quality data The collected indicators include pH value, turbidity, hardness, chloride ion concentration, iron ion concentration, and total bacterial count.

[0021] Data Acquisition Equipment and Installation: A high-precision online pH sensor (model: PH8000, measurement range 0-14pH, accuracy ±0.01pH) is used and installed in the return water pipe of the central air conditioning water system, at least 3 times the pipe diameter away from the pipe bend to ensure stable measurement; a turbidity sensor (model: TU5000, measurement range 0-100NTU, accuracy ±1%) is installed at the outlet of the circulating water pump; a hardness sensor (model: HD6000, measurement range 0-1000mg / L as CaCO3, accuracy ±2%) and a chloride ion sensor (model: CL7000, measurement range 0-2000mg / L, accuracy ±2%) are all installed on the main circulation pipe, avoiding areas with turbulent water flow.

[0022] Data collection frequency: Once every 10 seconds.

[0023] Environmental parameters Data collected: indoor temperature, indoor humidity, outdoor temperature, outdoor humidity, and atmospheric pressure.

[0024] Data Acquisition Equipment and Installation: Indoor temperature and humidity sensors (model: TH4000, temperature measurement range -40-80℃, accuracy ±0.5℃; humidity measurement range 0-100%RH, accuracy ±3%RH) are installed in different indoor areas, such as office areas and conference rooms, at a height of 1.5 meters above the ground; outdoor temperature, humidity, and atmospheric pressure sensors (model: WTHP6000, temperature measurement range -40-80℃, accuracy ±0.5℃; humidity measurement range 0-100%RH, accuracy ±3%RH; atmospheric pressure measurement range 300-1100hPa, accuracy ±1hPa) are installed in the ventilation area of ​​the building's roof.

[0025] Data collection frequency: Once every 30 seconds.

[0026] Air conditioning operating status data Data collected include: compressor power, fan speed, water flow rate, inlet and outlet water temperature, condenser pressure, and evaporator pressure.

[0027] Data Acquisition Equipment and Installation: Compressor power is acquired using a power sensor (model: PW3000, measurement range 0-500kW, accuracy ±0.5%); fan speed is acquired using a speed sensor (model: RS2000, measurement range 0-3000r / min, accuracy ±1r / min) mounted on the fan shaft; water flow is acquired using an electromagnetic flow meter (model: FM5000, measurement range 0-100m³ / min). 3 A temperature sensor (model: T6000, measuring range -50-150℃, accuracy ±0.1℃) is installed on the main water pipe; the inlet and outlet water temperatures are measured using a temperature sensor (model: P8000, measuring range 0-2MPa, accuracy ±0.2%FS) installed at the inlet and outlet pipes; the condenser and evaporator pressures are measured using a pressure sensor (model: P8000, measuring range 0-2MPa, accuracy ±0.2%FS) installed on the corresponding equipment.

[0028] Data collection frequency: Once every 5 seconds.

[0029] (ii) Data preprocessing Missing value handling For short-term, small-scale missing data (missing time less than 5 minutes), linear interpolation is used to fill the missing data. Based on the valid data points before and after the missing data, the missing values ​​are obtained through linear fitting.

[0030] For long-term missing or large-scale continuous missing (missing time greater than or equal to 5 minutes), the average value imputation method based on similar days is adopted. Historical dates with similar meteorological conditions and air conditioning operation modes to the current date are selected, and the average value of the corresponding time period is taken as the imputation value.

[0031] Outlier handling Outliers are detected using box plots. The quartiles (Q1, Q3) of the data are calculated, and the boundaries of outliers are determined as Q1-1.5×IQR and Q3+1.5×IQR (IQR is the interquartile range, i.e., Q3-Q1). Data exceeding these ranges are considered outliers.

[0032] For outliers, if they are caused by obvious sensor malfunctions, a moving average of the preceding and following data is used for correction; if they are caused by sudden interference, they are directly removed and interpolated using adjacent valid data.

[0033] Data standardization The min-max normalization method was used to unify the data to the [0,1] interval.

[0034] Multi-dimensional data construction The processed water quality data, environmental parameters, and air conditioning operation status data are aligned according to time series to form a multi-dimensional data matrix indexed by time. Each time point corresponds to a row of dynamic feature vectors, containing the values ​​of all monitoring indicators.

[0035] Step 102: Extract short-term local features through the gating mechanism of a multi-layer LSTM, capture long-term dependencies using the multi-head self-attention mechanism of Transformer, and establish an LSTM+Transformer hybrid network model. Specifically, in this embodiment, the input sensor time series data is used to extract short-term local features through the forget gate, input gate, and output gate of a multi-layer LSTM. By using a forget gate to filter out irrelevant noise, the output gate generates a hidden state sequence and preserves the short-term fluctuation patterns of water quality parameters.

[0036] Input the hidden state sequence output by LSTM into Transformer, and capture long-term dependencies through multi-head self-attention mechanism; The correlation between different time steps is quantified by self-attention weights, temporal sequence information is injected by position encoding, and local features of LSTM and global features of Transformer are weighted and fused to generate spatiotemporal feature vectors.

[0037] (I) LSTM Network Structure and Function Overall Architecture It adopts a three-layer progressive structure, with each layer using a gating mechanism to selectively process features, compressing feature dimensions layer by layer to extract core information.

[0038] First layer LSTM The system receives pre-processed multi-dimensional data (including time-series data on water quality, environment, and operational status) and performs initial screening through the collaborative action of input gates, forget gates, and output gates. The input gate determines which information in the new data is worth retaining; for example, it enhances the capture of a sudden increase in chloride ion concentration in the water. The forget gate filters out irrelevant historical noise, such as small short-term fluctuations. The output gate passes the processed, effective information to the next layer, outputting a 128-dimensional feature vector that focuses on capturing short-term patterns at the minute level, such as rapid fluctuations in turbidity within one hour.

[0039] Second layer LSTM Using the 128-dimensional features output from the first layer as input, the feature correlation is further enhanced. By optimizing the gating threshold, the focus is on the coupling relationship between equipment operating status and water quality changes, such as identifying short-term linkage features like increased turbidity after an increase in pump speed, and outputting a 64-dimensional feature vector to improve the accuracy of characterizing local dynamics.

[0040] Third layer LSTM By integrating the outputs of the first two layers, and by adjusting the gating sensitivity, redundant information is filtered out and key features are retained, a 32-dimensional short-term local feature vector is finally output. This vector reflects the core dynamic patterns in the past 3 hours, such as the short-term correlation pattern between water quality indicators and ambient temperature.

[0041] (II) Transformer Network Structure and Function Overall Architecture It consists of two cascaded encoder layers, each composed of a multi-head self-attention mechanism and a feedforward neural network, focusing on mining dependencies in long-term data.

[0042] Multi-head self-attention mechanism The 32-dimensional feature vector output by the LSTM is split into eight parallel subspaces for processing. Each subspace independently calculates the correlation weights of data from different locations. For example, one subspace focuses on the correlation between water hardness at 9 AM daily and operating pressure at 10 PM the previous day, while another subspace focuses on the relationship between total bacterial count on Monday and ambient humidity on Saturday of the previous week. The results from the eight subspaces are then integrated to form a comprehensive judgment on long-term dependencies, effectively capturing periodic patterns across days and weeks.

[0043] Feedforward Neural Network The features output by multi-head attention are subjected to a nonlinear transformation, and the feature representation is enhanced through two progressive layers. The first layer temporarily expands the feature dimension to accommodate more details. After activation processing, the second layer restores the dimension, which preserves key information and avoids feature redundancy, ensuring that the output 32-dimensional feature vector contains both short-term fluctuations and long-term trends.

[0044] (III) Hybrid Network Collaborative Operation Process The short-term local feature vector output by the LSTM is directly input into the Transformer network. The former provides fine dynamics at the minute to hour level, while the latter extracts cross-time-period correlation patterns from it. For example, when the LSTM captures a short-term feature of a rapid drop in pH value within a 10-minute period, the Transformer combines it with data from the same time period over the past 7 days to determine whether the drop is an occasional fluctuation or a precursor to a periodic fault. Finally, it outputs a comprehensive vector that integrates spatiotemporal features, containing both real-time dynamic details and long-term variation patterns.

[0045] Step 103: Optimize the hyperparameters of the hybrid network model using the improved IWOA whale optimization algorithm to obtain the target hybrid network model. Use the target hybrid network model to predict multi-dimensional data and output spatiotemporal feature vectors. Specifically, in this embodiment, a certain number of hyperparameter combinations are randomly generated as the initial population, with each individual representing a set of hyperparameters; each set of hyperparameters is substituted into the LSTM+Transformer hybrid network model, and the prediction error of the model on the training set is used as the fitness value. The population is updated using a search strategy based on the improved IWOA whale optimization algorithm, including predation, encirclement, and search operations, to find a better combination of hyperparameters. The optimal combination of hyperparameters obtained when the number of iterations reaches the preset maximum value is the target hyperparameter. Substituting the target hyperparameter into the model yields the target hybrid network model.

[0046] Specifically, (I) Details of core algorithm improvements Adaptive weight adjustment mechanism The search range is dynamically adjusted according to the iteration progress: In the initial stage (the first 30% of iterations), a larger weight is used to allow the algorithm to explore a wide range of hyperparameters, such as simultaneously testing combinations of different numbers of LSTM layers (2-5 layers) and different numbers of Transformer heads (4-16 heads); as the iteration progresses, the weight gradually decreases, and the search range focuses on the area around the current optimal solution. For example, when the LSTM is initially determined to be 3 layers, only the number of hidden units in each layer is fine-tuned (within ±20) to balance global exploration and local refinement.

[0047] Simulated annealing fusion strategy During the iteration process, when the performance of a new solution is slightly worse than the current best solution, it is not discarded directly. Instead, the acceptance probability is set according to the iteration stage: the acceptance probability is higher in the early stage (20%), allowing potential high-quality solutions to be retained to avoid getting trapped in local optima. For example, a set of hyperparameters may have a slightly larger error in the early stage of training, but may perform more stably in the later stage; the acceptance probability is reduced in the later stage (5%) to ensure that the algorithm converges quickly to a stable solution.

[0048] (II) Hyperparameter Optimization Process Population initialization 100 sets of random hyperparameter combinations are generated, covering the learning rate range (0.0001-0.01), number of hidden units (32-256), number of layers (2-5) of LSTM, as well as key parameters such as the number of heads (4-16) and number of encoder layers (1-3) of Transformer, to ensure the comprehensiveness of the initial search.

[0049] fitness assessment criteria Each set of hyperparameters is substituted into the hybrid network, and the model's performance in predicting historical data (the prediction error of water quality indicators) is used to determine its fitness; the smaller the error, the higher the fitness of that set of hyperparameters. Simultaneously, a generalization ability assessment is introduced, prioritizing combinations that perform well on the validation set to avoid overfitting.

[0050] Iterative optimization process In each iteration, the algorithm updates the population by simulating whale predation behavior: local fine-tuning (fine-tuning the learning rate) is performed on hyperparameter combinations with high fitness, and large-scale mutation (changing the number of Transformer heads) is performed on combinations with low fitness.

[0051] By combining adaptive weights and simulated annealing strategies, the optimization of the top 20% of high-quality combinations is focused after the 50th iteration. After the 80th iteration, the best 5 combinations are locked for fine-tuning. Finally, after 100 iterations, the hyperparameter combination with the best overall performance is output.

[0052] Step 104: Construct a GNN graph neural network to abstract the water system of the central air conditioning system into a topological graph of nodes and edges, and model the dynamic coupling relationship between the equipment. Specifically, in this embodiment, the initial attribute features of the receiving node include the encoded values ​​of basic attributes and the standardized values ​​of real-time state attributes, and each node is transformed into a 128-dimensional initial feature vector. Two graph convolutional layers are set up to aggregate node information. The first layer focuses on the information interaction between directly adjacent nodes, and the second layer expands the aggregation scope to include the information of indirectly connected nodes. A non-linear activation function is introduced after each graph convolution to enhance the model's expressive power and to process the aggregated features.

[0053] Specifically, (I) Abstraction of the topology of the central air conditioning water system Node definition and attributes Equipment nodes: These encompass core equipment such as circulating water pumps, heat exchangers, filters, cooling towers, water supply valves, and drain valves. Each equipment node includes basic attributes and real-time status attributes. Basic attributes include equipment model, rated power, design flow rate, and installation location. For example, a circulating water pump node might be labeled as model ISG100-200 with a rated flow rate of 50 m³ / h. 3 / h, installed in an underground machine room; real-time status attributes include current operating power, inlet and outlet pressure, fault codes, etc., such as the filter node updating the current differential pressure (reflecting the degree of blockage) and cumulative running time in real time.

[0054] Pipe nodes: Pipe components such as straight pipe sections, elbows, and tees in a water system are abstracted as nodes. Basic attributes include pipe diameter, material, length, and slope, such as a DN100 carbon steel straight pipe section with a length of 5 meters; real-time status attributes include flow velocity and average temperature in the pipe, such as real-time monitoring of local resistance coefficient changes for elbow nodes.

[0055] Sensor Nodes: These correspond to various online monitoring sensors. Their basic attributes include sensor model, measurement range, and installation location. Real-time status attributes include current measurement value, calibration time, and signal stability. For example, a turbidity sensor node may be labeled with the current measurement value of 15 NTU and the last calibration time.

[0056] Edge definition and properties Physical connection edge: Used to indicate the actual connection relationship between nodes, such as the connection between the outlet of a circulating water pump and the inlet of a filter through a pipe. The attributes of the edge include the connection method (flange connection, welding, etc.), sealing type, maximum pressure, etc. For example, an edge is marked as a flange connection, the sealing material is nitrile rubber, and the pressure resistance is 1.6MPa.

[0057] Functional association edges: These reflect the operational influence relationship between nodes. For example, there is a water level regulation relationship between the water supply valve and the expansion tank. The attributes of the edge are the influence weight (between 0 and 1) and the response delay time. When the opening of the water supply valve changes, the response delay time of the water level in the expansion tank is 30 seconds, and the influence weight is set to 0.8 (indicating a strong association).

[0058] (II) Design of GNN Graph Neural Network Structure Input layer: Receives the initial attribute features of the nodes, including the encoded values ​​of basic attributes (device model converted to classification code) and the standardized values ​​of real-time status attributes (stress values ​​converted to relative values ​​in the 0-1 range). Each node is transformed into a 128-dimensional initial feature vector.

[0059] Graph Convolutional Layers: Two graph convolutional layers are used to aggregate node information. The first layer focuses on the information interaction between directly adjacent nodes, such as aggregating the status of connected pipes and valves for a water pump node. The second layer expands the aggregation scope to include information from indirectly connected nodes, such as a cooling tower node acquiring temperature data from a distant heat exchanger. Each convolutional layer uses a weight matrix to weight and fuse the features of neighboring nodes, with the weights dynamically allocated based on the strength of the edge association.

[0060] Activation layer: A non-linear activation function is used to enhance the model's expressive power. It is introduced after each graph convolution layer to process the aggregated features and highlight key information (feature signals of abnormal states).

[0061] Output layer: Outputs the updated feature vector of each node, maintaining a dimension of 128. The features include the node's own state and its association with the entire system, providing a basis for generating subsequent control schemes.

[0062] (III) Modeling of dynamic coupling relationships between devices Dynamic updates to association strength: The weight values ​​of edges are adjusted based on real-time operational data. When the circulating water pump speed increases, the weight of its connection edge with the downstream pipeline increases (indicating enhanced water flow influence); when the filter pressure difference exceeds the threshold (clogging), the weight of its connection edge with the inlet pipeline decreases (indicating flow attenuation). The weight update cycle is synchronized with the data acquisition cycle (minimum 5 seconds / time).

[0063] Fault correlation modeling: To address the cascading effects of equipment failures, fault propagation paths are marked in the graph. When a water pump stops, the pipe nodes directly connected to it are marked as experiencing a sudden decrease in flow. This state is propagated downstream to the heat exchanger through edges, adding a correlation feature indicating a decrease in heat exchange efficiency to its feature vector, helping the model identify the scope of the fault's impact.

[0064] Step 105: Use the spatiotemporal feature vector as the initial embedding of the nodes in the GNN graph neural network, aggregate the neighbor information of each node, update the node state through the GCN graph convolutional layer, and output the operation control scheme.

[0065] Specifically, in this embodiment, the GNN graph neural network outputs a device control scheme based on the final node state, which includes at least pump speed adjustment, valve opening adjustment, heat exchanger area adjustment, and filter replacement.

[0066] Specifically, Based on the final node state output by GCN, a scheme is generated by combining a preset control rule base. The rule base includes basic rules (the filter needs to be backwashed when the turbidity exceeds 20 NTU) and dynamic rules (if the risk of heat exchanger scaling is positively correlated with the water pump flow rate, the water pump speed is reduced first). The rules are dynamically adapted according to the coupling relationship of the equipment. For example, when both the cooling tower and the heat exchanger are abnormal, the cooling tower with a larger impact range is controlled first.

[0067] Equipment control content Circulating water pump: Adjust the speed according to the energy consumption-flow balance coefficient in the node status. Under the premise of meeting the end demand, if the coefficient shows high energy consumption and low flow, reduce the speed by 5%-10%; if the neighboring pipeline node indicates insufficient pressure, increase the speed by 3%-5%.

[0068] Valves: Adjust the opening degree according to the flow and pressure characteristics of the neighbors. For example, when the pressure of the inlet pipe is too high, close the corresponding valve by 10%-15%; when the return water temperature in a certain area is too high, open the valve of that branch by 5%-8%.

[0069] Filter: Backwashing or filter replacement prompts are triggered based on the clogging risk value. Backwashing is initiated immediately when the risk value exceeds 80% (lasts for 30 seconds). If the threshold is still exceeded after 3 backwashes, a filter replacement alarm is issued.

[0070] Cooling tower: Adjust the fan speed and water replenishment frequency based on the characteristics of ambient temperature and return water temperature. For example, when the outdoor temperature exceeds 35℃, increase the fan speed by 20% and increase water replenishment by 1 time per hour.

[0071] Water treatment device: When the chemical demand index in the node status increases, the dosing pump is automatically started to add corrosion inhibitor or bactericide at 0.5%-1% of the current water volume.

[0072] Its beneficial effects are as follows: 1. It can comprehensively consider the mutual influence between equipment, resulting in more scientific and coordinated control schemes. For example, when adjusting the water pump speed, it comprehensively considers the heat exchanger efficiency, filter clogging, etc., avoiding increased energy consumption caused by frequent adjustments of a single device. This improves resource utilization while enhancing control efficiency. 2. It effectively reduces equipment failures caused by water quality issues. Simultaneously, by optimizing equipment operating parameters, it ensures equipment operates under reasonable conditions, reducing wear and aging, extending equipment lifespan, extending replacement cycles, and lowering replacement costs and downtime. 3. The system can automatically generate equipment operation reports and control records, providing managers with comprehensive and accurate equipment operation information, facilitating data analysis and decision-making, and improving the level of management precision and work efficiency.

[0073] Please see Figure 2 In a method for intelligent water quality monitoring and operation control of a central air conditioning system, the hyperparameters of a hybrid network model are optimized using an improved IWOA whale optimization algorithm to obtain a target hybrid network model. The target hybrid network model is then used to predict multi-dimensional data and output spatiotemporal feature vectors. The steps include: Step 201: Randomly generate a certain number of hyperparameter combinations as the initial population, with each individual representing a set of hyperparameters; substitute each set of hyperparameters into the LSTM+Transformer hybrid network model, and use the prediction error of the model on the training set as the fitness value; Step 202: Update the population using the improved IWOA whale optimization algorithm search strategy, including predation, encirclement and search operations, to find a better combination of hyperparameters. Step 203: When the number of iterations reaches the preset maximum value, the optimal combination of hyperparameters is the target hyperparameters. Substitute the target hyperparameters into the model to obtain the target hybrid network model.

[0074] Please see Figure 3 In a central air conditioning system water quality intelligent monitoring and operation control device, the central air conditioning system water quality intelligent monitoring and operation control device includes the following modules: The multi-dimensional data acquisition module is used to collect water quality data, environmental parameters and operating status data of central air conditioning equipment in real time, and to preprocess the collected data to obtain multi-dimensional data. The hybrid model building module is used to extract short-term local features through the gating mechanism of a multi-layer LSTM and capture long-term dependencies using the multi-head self-attention mechanism of the Transformer, thus building an LSTM+Transformer hybrid network model. The feature vector generation module is used to optimize the hyperparameters of the hybrid network model using the improved IWOA whale optimization algorithm to obtain the target hybrid network model. The target hybrid network model is then used to predict multi-dimensional data and output spatiotemporal feature vectors. The graph neural network building module is used to construct a GNN graph neural network, which abstracts the water system of the central air conditioning system into a topological graph of nodes and edges, and models the dynamic coupling relationship between equipment. The regulation scheme generation module is used to use spatiotemporal feature vectors as the initial embedding of nodes in a GNN graph neural network, aggregate neighbor information for each node, update the node state through a GCN graph convolutional layer, and output the running regulation scheme.

[0075] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent monitoring and operation control of water quality in a central air conditioning system, characterized in that, The intelligent water quality monitoring and operation control method for the central air conditioning system includes the following steps: Real-time data collection of water quality, environmental parameters, and operating status data from central air conditioning equipment; preprocessing of the collected data to obtain multi-dimensional data. A hybrid LSTM+Transformer network model is established by extracting short-term local features through the gating mechanism of a multi-layer LSTM and capturing long-term dependencies using the multi-head self-attention mechanism of the Transformer. The hyperparameters of the hybrid network model are optimized using the improved IWOA whale optimization algorithm to obtain the target hybrid network model. The multidimensional data is then predicted using the target hybrid network model, and a spatiotemporal feature vector is output. A graph neural network (GNN) is constructed to abstract the water system of the central air conditioning system into a topological graph of nodes and edges, and to model the dynamic coupling relationship between the equipment. The spatiotemporal feature vector is used as the initial embedding of the nodes in the GNN graph neural network. The neighbor information of each node is aggregated, the node state is updated through the GCN graph convolutional layer, and the operation control scheme is output.

2. The method for intelligent monitoring and operation control of water quality in a central air conditioning system as described in claim 1, characterized in that, The system collects real-time water quality data, environmental parameters, and operational status data from the central air conditioning equipment. The collected data is preprocessed to obtain multi-dimensional data, including: Real-time data collection of water quality, environmental parameters, and operating status data of central air conditioning equipment; linear interpolation is used to fill in missing data to obtain cleaning data. Outliers in the cleaned data are detected using the box plot method. The quartiles of the data are calculated, and data that exceeds the range are identified as outliers and deleted to obtain the processed data. After normalizing the processed data, the water quality data, environmental parameters, and air conditioning operation status data are aligned according to the time series to obtain multi-dimensional data indexed by time.

3. The method for intelligent monitoring and operation control of water quality in a central air conditioning system as described in claim 1, characterized in that, The method involves extracting short-term local features through the gating mechanism of a multi-layer LSTM and capturing long-term dependencies using the multi-head self-attention mechanism of Transformer, thus establishing an LSTM+Transformer hybrid network model, including: Input sensor time-series data and extract short-term local features through the forget gate, input gate and output gate of a multi-layer LSTM; By using a forget gate to filter out irrelevant noise, the output gate generates a hidden state sequence and preserves the short-term fluctuation patterns of water quality parameters.

4. The method for intelligent monitoring and operation control of water quality in a central air conditioning system as described in claim 1, characterized in that, The method of extracting short-term local features through the gating mechanism of multi-layer LSTM and capturing long-term dependencies through the multi-head self-attention mechanism of Transformer to establish an LSTM+Transformer hybrid network model also includes: Input the hidden state sequence output by LSTM into Transformer, and capture long-term dependencies through multi-head self-attention mechanism; The correlation between different time steps is quantified by self-attention weights, temporal sequence information is injected by position encoding, and local features of LSTM and global features of Transformer are weighted and fused to generate spatiotemporal feature vectors.

5. The method for intelligent monitoring and operation control of water quality in a central air conditioning system as described in claim 1, characterized in that, The improved IWOA whale optimization algorithm is used to optimize the hyperparameters of the hybrid network model to obtain the target hybrid network model. This target hybrid network model is then used to predict the multi-dimensional data, outputting a spatiotemporal feature vector, including: A certain number of hyperparameter combinations are randomly generated as the initial population, with each individual representing a set of hyperparameters; each set of hyperparameters is substituted into the LSTM+Transformer hybrid network model, and the prediction error of the model on the training set is used as the fitness value. The population is updated using a search strategy based on the improved IWOA whale optimization algorithm, including predation, encirclement, and search operations, to find a better combination of hyperparameters. The optimal combination of hyperparameters obtained when the number of iterations reaches the preset maximum value is the target hyperparameter. Substituting the target hyperparameter into the model yields the target hybrid network model.

6. The method for intelligent monitoring and operation control of water quality in a central air conditioning system as described in claim 1, characterized in that, The construction of the GNN graph neural network abstracts the central air conditioning water system as a topological graph of nodes and edges, and models the dynamic coupling relationships between devices, including: The initial attribute features of the receiving nodes, including the encoded values ​​of basic attributes and the standardized values ​​of real-time state attributes, are transformed into a 128-dimensional initial feature vector for each node. Two graph convolutional layers are set up to aggregate node information. The first layer focuses on the information interaction between directly adjacent nodes, and the second layer expands the aggregation scope to include the information of indirectly connected nodes. A non-linear activation function is introduced after each graph convolution to enhance the model's expressive power and to process the aggregated features.

7. The method for intelligent monitoring and operation control of water quality in a central air conditioning system as described in claim 1, characterized in that, The process of using the spatiotemporal feature vector as the initial embedding of nodes in the GNN graph neural network, aggregating neighbor information for each node, updating node states through GCN graph convolutional layers, and outputting an operational control scheme includes: Based on the final node state, the GNN (Graph Neural Network) outputs a device control scheme, which includes at least pump speed adjustment, valve opening adjustment, heat exchanger area adjustment, and filter replacement.

8. A central air conditioning system water quality intelligent monitoring and operation control system, characterized in that, The intelligent water quality monitoring and operation control system for the central air conditioning system includes the following modules: The multi-dimensional data acquisition module is used to collect water quality data, environmental parameters and operating status data of central air conditioning equipment in real time, and to preprocess the collected data to obtain multi-dimensional data. The hybrid model building module is used to extract short-term local features through the gating mechanism of a multi-layer LSTM and capture long-term dependencies using the multi-head self-attention mechanism of the Transformer, thus building an LSTM+Transformer hybrid network model. The feature vector generation module is used to optimize the hyperparameters of the hybrid network model using the improved IWOA whale optimization algorithm to obtain the target hybrid network model. The target hybrid network model is then used to predict the multi-dimensional data and output spatiotemporal feature vectors. The graph neural network building module is used to construct a GNN graph neural network, which abstracts the water system of the central air conditioning system into a topological graph of nodes and edges, and models the dynamic coupling relationship between equipment. The regulation scheme generation module is used to use the spatiotemporal feature vector as the initial embedding of the nodes of the GNN graph neural network, aggregate the neighbor information of each node, update the node state through the GCN graph convolutional layer, and output the operation regulation scheme.

9. The intelligent water quality monitoring and operation control system for a central air conditioning system as described in claim 8, characterized in that, The graph neural network construction module includes the following sub-modules: The receiving submodule is used to receive the initial attribute features of the nodes, including the encoded values ​​of the basic attributes and the standardized values ​​of the real-time state attributes. Each node is transformed into a 128-dimensional initial feature vector. The configuration submodule is used to set up two graph convolutional layers to aggregate node information. The first layer focuses on the information interaction between directly adjacent nodes, and the second layer expands the aggregation scope to include information of indirectly connected nodes. A submodule is introduced to introduce a non-linear activation function after each graph convolution to enhance the model's expressive power and to process the aggregated features.

10. The intelligent water quality monitoring and operation control system for a central air conditioning system as described in claim 8, characterized in that, The control scheme generation module includes the following sub-modules: The output submodule is used by the GNN graph neural network to output equipment control schemes based on the final node state, including at least pump speed adjustment, valve opening adjustment, heat exchanger area adjustment, and filter replacement.

Images