A cooling tower adaptive temperature control system based on deep learning

Abstract

This invention discloses a deep learning-based adaptive temperature control system for cooling towers, comprising: a packing heat exchange path construction module for constructing a set of packing heat exchange paths; a water film trajectory acquisition module for generating a sequence of water film attachment trajectories; a tension topology construction module for generating a water film tension topology map; an environment-driven thermal potential construction module for constructing an environment-driven thermal potential map; a tension-thermal potential coupling module for generating a tension-thermal potential coupling structure diagram; a gas-liquid coupling state modeling module for calculating gas-liquid coupling state characteristic values ​​and constructing a gas-liquid coupling state sequence; a temperature trend prediction module for outputting future temperature change trends based on an improved deep Koopman network and incorporating a dynamic spectrum hierarchical evolution mechanism; and an operating parameter adjustment module for adjusting the operating parameters of the cooling tower. This invention employs tension-thermal potential coupling modeling and an improved deep Koopman network to achieve cooling tower temperature prediction and adaptive control.

Description

Technical Field

[0001] This invention relates to the field of cooling tower temperature control technology, and in particular to a deep learning-based adaptive temperature control system for cooling towers. Background Technology

[0002] With the increasing demands for energy efficiency control, temperature stability, and equipment reliability in industrial circulating cooling processes, adaptive temperature control technology for cooling towers under complex environmental conditions has received widespread attention. Existing cooling tower temperature regulation methods mainly rely on manual experience-based adjustment, fixed-rule control, or closed-loop control based on single temperature feedback to adjust operating parameters. However, these methods generally suffer from the following problems in practical applications: The heat exchange process in a cooling tower is simultaneously influenced by external ambient temperature, humidity, wind speed, and the distribution of the water film within the tower. Existing technologies typically rely solely on outlet water temperature, circulating water flow rate, or air flow rate for control, lacking a comprehensive characterization of the water film adhesion state, tension changes, and spatial distribution of environmental heat dissipation capacity on the packing surface. This results in insufficient characterization of the internal mechanisms of the heat exchange process and significant regulatory lag. The water film flow on the surface of the cooling tower packing exhibits significant spatiotemporal nonuniformity. Water film adhesion, confluence, breakage, and local stretching and contraction processes directly affect gas-liquid heat exchange efficiency. Existing technologies struggle to continuously model the water film flow trajectory and its tension topology, leading to low accuracy in heat exchange state identification and difficulty in timely detection of local instability areas. The gas-liquid coupling behavior during cooling tower operation exhibits significant nonlinearity, non-stationarity, and multi-timescale characteristics. Traditional PID control, rule-based threshold control, and ordinary neural network prediction methods typically lack the ability to jointly model spatial topological relationships, multi-path unit state evolution, and multi-frequency dynamic characteristics, easily resulting in unstable temperature prediction, large parameter fluctuations, and insufficient adaptability to environmental disturbances.

[0003] Therefore, how to provide a deep learning-based adaptive temperature control system for cooling towers is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a deep learning-based adaptive temperature control system for cooling towers. This invention employs tension-thermal potential coupling modeling and an improved deep Koopman network to perform multi-scale modeling and prediction of the gas-liquid coupling process in the cooling tower. This enables accurate prediction of the outlet water temperature change trend and adaptive adjustment of operating parameters, effectively improving temperature control accuracy and system stability, enhancing adaptability to complex environmental disturbances, reducing energy consumption, and improving operating efficiency.

[0005] According to an embodiment of the present invention, a deep learning-based adaptive temperature control system for cooling towers includes: The packing heat exchange path construction module is used to perform three-dimensional modeling of the cooling tower packing structure, generate heat exchange path units, and construct a packing heat exchange path set. The water film trajectory acquisition module is used to continuously acquire the flow state of circulating water on the surface of the packing material based on the heat exchange path set of the packing material, and generate a water film adhesion trajectory sequence. The tension topology construction module is used to spatially divide the packing surface according to the water film attachment trajectory sequence, identify the water film stable region, water film stretching region and water film contraction region in each heat exchange path unit, and generate a water film tension topology map. The environment-driven thermal potential construction module is used to collect external environmental parameters of the cooling tower, divide the environmental areas and calculate the heat dissipation capacity value of each environmental area, and construct an environment-driven thermal potential map. The tension-thermal potential coupling module is used to spatially map the water film tension topology map with the environmental driven thermal potential map, and to map the heat dissipation capacity of the environmental area to the corresponding heat exchange path unit, generating a tension-thermal potential coupling structure map. The gas-liquid coupling state modeling module is used to calculate the gas-liquid coupling state characteristic values ​​of each heat transfer path unit at each sampling time based on the tension-thermal potential coupling structure diagram, and construct the gas-liquid coupling state sequence. The temperature trend prediction module is used to input the tension-thermal potential coupling structure diagram, gas-liquid coupling state sequence and historical effluent temperature data into the improved deep Koopman network, and process them sequentially through the state encoding unit, dynamic feature mapping unit, spectral evolution analysis unit, Koopman state evolution unit and prediction output unit. The spectral evolution analysis unit introduces a dynamic spectral layer evolution mechanism to output the future temperature change trend. The operating parameter adjustment module is used to adjust the operating parameters of the cooling tower according to future temperature change trends, so as to achieve adaptive control of the cooling tower outlet water temperature.

[0006] Optionally, the packing heat exchange path construction module specifically comprises: Obtain the structural parameters of the cooling tower packing structure, including the packing layer thickness, layer spacing, and layer tilt angle, and establish a three-dimensional coordinate model of the packing structure based on the structural parameters, and establish a three-dimensional rectangular coordinate system with the bottom of the packing as the coordinate origin; The surface of the filler sheet in the three-dimensional coordinate model is discretized and divided into multiple surface mesh units according to a preset spatial resolution. Each surface mesh unit corresponds to a unique three-dimensional coordinate. In each surface grid cell, a gravity direction vector and a sheet tilt angle direction vector are established, and the vectors are superimposed to obtain the water film flow direction vector; Between adjacent surface grid cells, when the angle between two water film flow direction vectors is less than a preset angle threshold, the adjacent surface grid cells are connected sequentially along the water film flow direction to form a water film adhesion path. In the three-dimensional coordinate model, an airflow direction vector field along the air inlet direction is established, and the channels between the packing sheets are divided into multiple volumetric grid units. Between adjacent volume grid cells, when the angle between the airflow direction vectors is less than a preset angle threshold, the adjacent volume grid cells are connected sequentially along the airflow direction to form an airflow path. In the surface grid cells where the water film adhesion path and the air flow path intersect, surface grid cells with a water film coverage rate greater than a preset coverage rate threshold and an air flow velocity greater than a preset flow velocity threshold are selected as water-air contact areas, and the corresponding three-dimensional coordinates are recorded. Centered on the water-air contact area, adjacent surface grid units that satisfy the condition that the angle between the water film flow direction vectors is less than a preset angle threshold and that the air channels are connected are aggregated to form a heat exchange path unit. Each heat exchange path unit encloses at least one water-air contact area. Each heat exchange path unit is numbered and its corresponding three-dimensional coordinates are recorded to generate a set of packing heat exchange paths.

[0007] Optionally, the water film trajectory acquisition module specifically comprises: Read the unit number and corresponding three-dimensional coordinates of each heat exchange path in the packing heat exchange path set, and establish an index relationship between the surface mesh units corresponding to each heat exchange path unit; When the cooling tower inlet water enters the packing structure area, the surface grid unit is continuously monitored. When the water film coverage rate changes from less than the preset coverage rate threshold to equal to or greater than the preset coverage rate threshold in the surface grid unit, the corresponding surface grid unit is marked as the water film attachment start position, and the corresponding heat exchange path unit number and timestamp are recorded. According to the preset time sampling interval, the surface grid unit is periodically sampled, and the water film coverage and water film flow direction vector of the surface grid unit are obtained at each sampling time. Between adjacent sampling times, within the same heat exchange path unit, when the water film coverage of adjacent surface grid units is greater than a preset coverage threshold and the angle between the corresponding water film flow direction vectors is less than a preset angle threshold, the adjacent surface grid units are connected sequentially according to the water film flow direction to form a water film flow trajectory segment. The water film flow trajectory segments formed at continuous sampling times are spliced ​​together in chronological order, and the connection relationship across heat exchange path units is established by combining the heat exchange path unit number to form the flow trajectory of the water film along the heat exchange path unit. At the same sampling time, when multiple water film flow trajectory segments converge in the same surface grid cell and the water film coverage is greater than a preset multiple of the average coverage of adjacent surface grid cells, the corresponding surface grid cell is marked as the water film confluence location, and the corresponding heat exchange path cell number and timestamp are recorded. Between adjacent sampling times, when the water film coverage of the surface grid unit changes from greater than the preset coverage threshold to less than the preset fracture threshold, the corresponding surface grid unit is marked as the water film fracture location, and the corresponding heat exchange path unit number and timestamp are recorded. By integrating the water film attachment start position, the water film flow trajectory along the heat exchange path unit, the water film confluence position, and the water film break position in chronological order, a water film attachment trajectory sequence is generated.

[0008] Optionally, the tension topology construction module specifically comprises: Read the water film attachment trajectory sequence and integrate it in chronological order: the water film attachment start position, the water film flow trajectory along the heat exchange path unit, the water film confluence position, and the water film break position. Using the initial position of water film adhesion as the starting point, the corresponding surface grid units are sequentially expanded based on the flow trajectory of the water film along the heat transfer path unit to construct the temporal distribution chain of the surface grid units within each heat transfer path unit. Within the same heat exchange path unit, the changes in water film coverage and the changes in the angle between water film flow direction vectors of each surface grid unit in the time-series distribution chain are calculated between adjacent sampling times. When the change in water film coverage of a surface grid cell between several consecutive sampling times is less than or equal to a preset stability threshold, and the change in the angle between the water film flow direction vectors is less than or equal to a preset change threshold, and the surface grid cell does not belong to the water film confluence location or the water film breakage location, the corresponding surface grid cell is marked as a water film stable region. When a surface grid cell is located on the flow trajectory of the water film along the heat transfer path cell, and the increase in water film coverage between several consecutive sampling times is greater than a preset stability threshold, or the surface grid cell is adjacent to the water film confluence position and the change in the angle between the water film flow direction vectors is greater than a preset change threshold, the corresponding surface grid cell is marked as a water film stretching region. When a surface grid cell is located at the end of the flow trajectory of the water film along the heat transfer path cell, and the reduction in water film coverage between several consecutive sampling times is greater than a preset stability threshold, or when the surface grid cell is adjacent to the water film breakage position and the change in the angle between the water film flow direction vectors is greater than a preset change threshold, the corresponding surface grid cell is marked as a water film contraction region. Within the same heat exchange path unit, spatially adjacent and identically labeled surface mesh units are aggregated to form tension nodes, and the heat exchange path unit number, three-dimensional coordinates, and label type corresponding to each tension node are recorded. Based on the chronological order from the initial position of water film attachment to the water film confluence position, and from the water film confluence position to the water film breakage position, a directed connection relationship is established between tension nodes to generate a water film tension topology map.

[0009] Optionally, the environment-driven thermal potential construction module specifically comprises: External environmental parameters are collected on the air inlet side of the cooling tower, including ambient temperature, ambient humidity, and ambient wind speed. In the three-dimensional coordinate model, the plane where the air inlet of the cooling tower is located is selected as the air inlet section, and it is divided into multiple environmental areas according to the preset grid size. Each environmental area is numbered and its corresponding spatial coordinates are recorded. The ambient temperature, ambient humidity, and ambient wind speed are assigned to the corresponding environmental area. When there are multiple sampling values ​​in the same environmental area, the arithmetic mean of each sampling value is used to obtain the ambient area temperature, ambient area humidity, and ambient area wind speed. The ambient temperature and ambient humidity are converted to obtain the ambient wet-bulb temperature. The temperature difference of the environmental area is obtained by subtracting the wet-bulb temperature of the environmental area from the temperature of the environmental area, and the air volume flow rate is obtained by multiplying the wind speed of the environmental area by the area of ​​the environmental area. The heat dissipation capacity of the environmental area is obtained by multiplying the temperature difference of the environmental area by the air volume flow rate. Arrange the heat dissipation capacity values ​​of each environmental area according to spatial location to construct an environmental-driven thermal potential map.

[0010] Optionally, the tension-thermal potential coupling module specifically comprises: Based on the tension nodes formed by the aggregation of multiple surface mesh units in the water film tension topology diagram, the heat transfer path unit number, three-dimensional coordinates and label type corresponding to each tension node are read, and the directed connection relationship between tension nodes is obtained according to the time sequence from the water film attachment start position to the water film confluence position, and from the water film confluence position to the water film breakage position. Read the spatial coordinates and heat dissipation capacity values ​​of each environmental region from the environmental-driven thermal potential map; For each tension node, all surface mesh elements in the tension node are extracted one by one, and the three-dimensional coordinates of each surface mesh element are projected along the air inlet direction onto the plane where the air inlet section is located to obtain the projected coordinates; Based on the position of the projected coordinates in the air inlet section, the corresponding environmental area is determined, and the heat dissipation capacity value of the environmental area is read. The heat dissipation capacity values ​​corresponding to all surface mesh elements within the tension node are accumulated and averaged to obtain the equivalent heat dissipation capacity value of the corresponding tension node. The equivalent heat dissipation capacity value is associated with the label type of the corresponding tension node, and the tension-thermal potential coupling structure diagram is generated by combining the directed connection relationship between the tension nodes.

[0011] Optionally, the gas-liquid coupling state modeling module specifically comprises: Based on the tension-thermal potential coupling structure diagram, the heat transfer path unit number, equivalent heat dissipation capacity value and marking type corresponding to each tension node are read; The tension nodes of each heat exchange path unit are classified according to the marking type of the tension nodes. The marking type includes water film stable region, water film stretching region and water film contraction region. At each sampling time, the equivalent heat dissipation capacity values ​​of the tension nodes corresponding to the stable region of the water film in each heat exchange path unit are accumulated to obtain the stable heat dissipation term; the equivalent heat dissipation capacity values ​​of the tension nodes corresponding to the stretched region of the water film are accumulated to obtain the stretched heat dissipation term; and the equivalent heat dissipation capacity values ​​of the tension nodes corresponding to the contracted region of the water film are accumulated to obtain the contracted heat dissipation term. Multiply the stretching heat dissipation term by a set enhancement coefficient, multiply the contraction heat dissipation term by a set weakening coefficient, and add the enhanced stretching heat dissipation term, the weakened contraction heat dissipation term, and the stable heat dissipation term to obtain the gas-liquid coupling state characteristic value of each heat exchange path unit at the current sampling time. At each sampling moment, the gas-liquid coupling state feature values ​​corresponding to each heat exchange path unit are spliced ​​together according to the heat exchange path unit number order to obtain the path state vector at the current sampling moment. The path state vectors obtained at different sampling times are arranged in chronological order to generate a gas-liquid coupling state sequence.

[0012] Optionally, the temperature trend prediction module specifically comprises: The tension-thermal potential coupling structure diagram, gas-liquid coupling state sequence and historical outlet water temperature data are input into the state encoding unit. The equivalent heat dissipation capacity value, label type and heat exchange path unit number corresponding to each tension node are extracted according to the tension node order. The label type is numerically encoded to form a node feature vector. Align the path state vector in the gas-liquid coupling state sequence with the historical effluent temperature data according to the sampling time, and connect the path state vector at the same sampling time with the corresponding effluent temperature to form a time feature vector; match the node feature vector of each tension node with the time feature vector at the corresponding sampling time and connect them to generate the input feature sequence. The input feature sequence is fed into the dynamic feature mapping unit. Based on the directed connection relationship between tension nodes, each tension node and its adjacent tension nodes with directed connection relationship are taken as neighborhood nodes. A three-layer graph convolution structure is used for feature propagation. The adjacency range of each graph convolution layer is set to the first-order neighborhood, the convolution kernel size is set to 3×3, and the number of output channels is set to 32, 64 and 64 respectively. Linear transformation and ReLU nonlinear activation processing are performed on the output of each layer to generate dynamic features. The dynamic features are input into the spectral evolution analysis unit, and the dynamic spectral layer evolution mechanism is executed to generate spectral coupling features; The spectral coupling features are input into the Koopman state evolution unit, arranged in the order of sampling time to form a state feature sequence, and the spectral coupling features corresponding to each sampling time are used as the state vector. The predicted state vector for the next sampling time is obtained by multiplying the state vector at the current sampling time with the state evolution operator. The predicted state vector is then appended to the end of the state vector sequence in chronological order. At the same time, the state vector corresponding to the earliest sampling time is removed to form an updated state vector sequence. Based on the updated state vector sequence, the multiplication operation between the state vector and the state evolution operator is repeatedly performed to realize the step-by-step recursion of the state vector sequence and generate the state feature sequence at multiple future sampling times; The state feature sequences of multiple future sampling times are input into the prediction output unit, which consists of a fully connected layer and an output mapping layer. The fully connected layer performs feature mapping on the state feature sequences of multiple future sampling times, and the number of neurons in the fully connected layer is set to 64. The output mapping layer performs a linear transformation on the output of the fully connected layer to obtain the predicted value of the effluent temperature at the corresponding sampling time. The predicted values ​​of the effluent temperature at multiple sampling times are arranged in chronological order to output the future temperature change trend.

[0013] Optionally, the dynamic spectrum hierarchical evolution mechanism is specifically as follows: The difference between the dynamic features at adjacent sampling times is calculated. The dynamic features at the next sampling time are subtracted from the dynamic features at the previous sampling time. The absolute value of the difference result is processed. The absolute value results of each component in the same dynamic feature are accumulated and divided by the number of components to obtain the feature change of the corresponding dynamic feature. When the feature change is less than or equal to the set low-frequency threshold, the corresponding dynamic feature is classified as a low-frequency layer; when the feature change is greater than the set low-frequency threshold and less than the set high-frequency threshold, the corresponding dynamic feature is classified as a mid-frequency layer; when the feature change is greater than or equal to the set high-frequency threshold, the corresponding dynamic feature is classified as a high-frequency layer. The dynamic features of the low-frequency layer are accumulated time-by-time in chronological order, and the accumulated result is divided by the number of sampling times involved in the accumulation to generate low-frequency stationary features; The difference vector between adjacent sampling times is calculated sequentially for the dynamic features of the mid-frequency layer, where each difference vector is the dynamic feature of the next sampling time minus the dynamic feature of the previous sampling time. The sampling times are numbered in chronological order. The minimum number is subtracted from the number of each sampling time, and then the result is divided by the difference between the maximum and minimum numbers to obtain the normalized time weight. Each difference vector is multiplied by its corresponding normalized time weight and then summed. The summation result is divided by the sum of all normalized time weights to obtain the mid-frequency variation characteristics. The high-frequency layer dynamic features are processed in chronological order. First, the high-frequency layer dynamic features at all sampling times are accumulated component by component and divided by the number of sampling times to obtain the high-frequency layer average features. The deviation features are obtained by subtracting the average features of the high-frequency layer from the dynamic features of the high-frequency layer at each sampling time. By performing component-by-component absolute value processing on the deviation characteristics, the high-frequency amplitude characteristics corresponding to each sampling time are obtained; The high-frequency amplitude features corresponding to all sampling times are accumulated component by component in chronological order, and the accumulated result is divided by the number of sampling times to generate high-frequency disturbance features. The low-frequency stationary characteristics and the mid-frequency variation characteristics are added component by component to generate the mid-frequency coupling characteristics; The mid-frequency variation characteristics and high-frequency disturbance characteristics are added component by component to generate high-frequency coupling characteristics; The low-frequency stable features, mid-frequency coupling features, and high-frequency coupling features are spliced ​​together in order of frequency hierarchy to generate spectral coupling features.

[0014] Optionally, the operating parameter adjustment module specifically comprises: Read future temperature change trends and obtain predicted effluent temperatures at multiple sampling times; The predicted effluent temperatures at multiple sampling times are summed and divided by the number of sampling times to obtain the average predicted temperature. The set target effluent temperature is then subtracted from the average predicted temperature to obtain the temperature deviation value. The operating parameters of the cooling tower are adjusted based on the temperature deviation value. These operating parameters include circulating water flow rate and air flow rate. When the temperature deviation exceeds the set positive threshold, increase the circulating water flow rate and air flow rate; When the temperature deviation is less than the set negative threshold, reduce the circulating water flow rate and air flow rate; The cooling tower is operated and controlled based on the adjusted circulating water flow and air flow to achieve adaptive adjustment of the outlet water temperature.

[0015] The beneficial effects of this invention are: This invention combines spatial coupling modeling of water film tension topology and environmentally driven thermal potential maps with temporal representation of gas-liquid coupling state sequences. It also introduces an improved deep Koopman network to address the problem of cooling tower heat exchange processes being affected by both uneven water film distribution and environmental disturbances. By using a set of packing heat exchange paths and a sequence of water film attachment trajectories, it achieves a continuous characterization of the water film flow process. In the tension topology construction module, tension nodes and directed connections are formed, allowing for a structured representation of the water film's stable, stretched, and contracted regions. Furthermore, in the tension-thermal potential coupling module, the equivalent heat dissipation capacity value is correlated with the tension nodes, achieving a unified description of spatial heat exchange capacity and water film state. Further, in the gas-liquid coupling state modeling module, different label types are used... The equivalent heat dissipation capacity of the tension nodes is accumulated and weighted to construct a gas-liquid coupling state sequence, enabling a time-series expression of the dynamic heat transfer capacity of the heat transfer path unit. In the temperature trend prediction module, by improving the synergistic effect of the state coding unit, dynamic feature mapping unit, spectral evolution analysis unit, Koopman state evolution unit, and prediction output unit of the deep Koopman network, a dynamic spectral hierarchical evolution mechanism is introduced to perform hierarchical processing and coupled expression of features at different time scales, thereby improving the modeling ability of complex gas-liquid coupling processes. Finally, the adaptive adjustment of circulating water flow and air flow is achieved through the operating parameter adjustment module, thereby improving the stability, prediction accuracy, and adaptability to changes in operating conditions of the cooling tower outlet water temperature control. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the structure of a deep learning-based adaptive temperature control system for cooling towers proposed in this invention. Figure 2 This is a schematic diagram of an improved deep Koopman network structure for a deep learning-based adaptive temperature control system for cooling towers proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1 and Figure 2 A deep learning-based adaptive temperature control system for cooling towers includes: The packing heat exchange path construction module is used to perform three-dimensional modeling of the cooling tower packing structure, generate heat exchange path units, and construct a packing heat exchange path set. The water film trajectory acquisition module is used to continuously acquire the flow state of circulating water on the surface of the packing material based on the heat exchange path set of the packing material, and generate a water film adhesion trajectory sequence. The tension topology construction module is used to spatially divide the packing surface according to the water film attachment trajectory sequence, identify the water film stable region, water film stretching region and water film contraction region in each heat exchange path unit, and generate a water film tension topology map. The environment-driven thermal potential construction module is used to collect external environmental parameters of the cooling tower, divide the environmental areas and calculate the heat dissipation capacity value of each environmental area, and construct an environment-driven thermal potential map. The tension-thermal potential coupling module is used to spatially map the water film tension topology map with the environmental driven thermal potential map, and to map the heat dissipation capacity of the environmental area to the corresponding heat exchange path unit, generating a tension-thermal potential coupling structure map. The gas-liquid coupling state modeling module is used to calculate the gas-liquid coupling state characteristic values ​​of each heat transfer path unit at each sampling time based on the tension-thermal potential coupling structure diagram, and construct the gas-liquid coupling state sequence. The temperature trend prediction module is used to input the tension-thermal potential coupling structure diagram, gas-liquid coupling state sequence and historical effluent temperature data into the improved deep Koopman network. The network is then processed sequentially through the state coding unit, dynamic feature mapping unit, spectral evolution analysis unit, Koopman state evolution unit and prediction output unit. The spectral evolution analysis unit introduces a dynamic spectral layer evolution mechanism to output the future temperature change trend. The operating parameter adjustment module is used to adjust the operating parameters of the cooling tower according to future temperature change trends, so as to achieve adaptive control of the cooling tower outlet water temperature.

[0019] In this embodiment, the packing heat exchange path construction module is specifically as follows: Obtain the structural parameters of the cooling tower packing structure, including the packing layer thickness, layer spacing, and layer tilt angle. Based on the structural parameters, establish a three-dimensional coordinate model of the packing structure and establish a three-dimensional rectangular coordinate system with the bottom of the packing as the origin. The surface of the filler sheet in the three-dimensional coordinate model is discretized and divided into multiple surface mesh units according to a preset spatial resolution. Each surface mesh unit corresponds to a unique three-dimensional coordinate. In each surface grid cell, a gravity direction vector and a sheet tilt angle direction vector are established, and the vectors are superimposed to obtain the water film flow direction vector; Between adjacent surface grid cells, when the angle between two water film flow direction vectors is less than a preset angle threshold, the adjacent surface grid cells are connected sequentially along the water film flow direction to form a water film adhesion path. In the three-dimensional coordinate model, an airflow direction vector field along the air inlet direction is established, and the channels between the packing sheets are divided into multiple volumetric grid units. Between adjacent volume grid cells, when the angle between the airflow direction vectors is less than a preset angle threshold, the adjacent volume grid cells are connected sequentially along the airflow direction to form an airflow path. In the surface grid cells where the water film adhesion path and the air flow path intersect, surface grid cells with a water film coverage rate greater than a preset coverage rate threshold and an air flow velocity greater than a preset flow velocity threshold are selected as water-air contact areas, and the corresponding three-dimensional coordinates are recorded. Centered on the water-air contact area, adjacent surface grid units that satisfy the condition that the angle between the water film flow direction vectors is less than a preset angle threshold and that the air channels are connected are aggregated to form a heat exchange path unit. Each heat exchange path unit encloses at least one water-air contact area. Each heat exchange path unit is numbered and its corresponding three-dimensional coordinates are recorded to generate a set of packing heat exchange paths; In the specific implementation process, the structural parameters of the cooling tower packing structure are first obtained, wherein the packing layer thickness is 0.5mm, the layer spacing is 20mm, and the layer tilt angle is 60°. Based on the above structural parameters, a three-dimensional coordinate model of the packing structure is established, and a three-dimensional rectangular coordinate system is established with the center point of the bottom of the packing as the coordinate origin. The X-axis is along the horizontal direction, the Y-axis is along the air inlet direction, and the Z-axis is along the vertical direction.

[0020] In the three-dimensional coordinate model, the surface of the filler sheet is discretized with a spatial resolution of 5mm×5mm. The filler sheet is divided into multiple surface mesh units, each of which corresponds to a unique three-dimensional coordinate point and is recorded as a node number.

[0021] In each surface grid cell, a gravity direction vector g=(0,0,-1) is established, and a sheet tilt angle direction vector is constructed based on the 60° tilt angle of the packing sheet. The gravity direction vector and the sheet tilt angle direction vector are vector superimposed to obtain the water film flow direction vector, which is used to characterize the flow trend of the water film on the surface grid cell.

[0022] Calculate the angle between the water film flow direction vectors between adjacent surface grid cells. When the angle is less than 20°, connect the adjacent surface grid cells sequentially according to the water film flow direction to form a continuous water film attachment path.

[0023] In the three-dimensional coordinate model, an airflow direction vector field is established along the Y-axis. The spatial region between the packing sheets is divided into volume grid cells with a size of 10mm×10mm×10mm. The angle between the airflow direction vectors of adjacent volume grid cells is calculated. When the angle is less than 20°, the volume grid cells are connected sequentially along the airflow direction to form an airflow path.

[0024] At the intersection of the water film attachment path and the air flow path, the corresponding surface grid cells are screened. The water film coverage is calculated by the ratio of the area occupied by the water film in a unit surface grid cell to the total area of ​​the grid, with a coverage threshold of 0.7. The air velocity is obtained by measuring the velocity in the volume grid cell, with a velocity threshold of 1.5 m / s. When both the water film coverage and the air velocity are greater than the coverage threshold, the corresponding surface grid cell is marked as a water-air contact area, and its three-dimensional coordinates are recorded.

[0025] Starting from the water-air contact area, adjacent surface grid cells are expanded and aggregated. During the expansion process, the angle between the water film flow direction vectors of adjacent surface grid cells is calculated. When the angle is less than 20° and the line connecting two surface grid cells does not spatially intersect with the packing sheet structure, the adjacent surface grid cells are included in the same set. The expansion continues until the condition is no longer met, forming heat transfer path cells. Each heat transfer path cell includes at least one water-air contact area. All heat transfer path cells are numbered, and the three-dimensional coordinates and spatial distribution relationship of all surface grid cells in each heat transfer path cell are recorded to construct a packing heat transfer path set for membrane tension analysis and heat transfer state modeling.

[0026] In this embodiment, the water film trajectory acquisition module specifically comprises: Read the unit number and corresponding three-dimensional coordinates of each heat exchange path in the packing heat exchange path set, and establish an index relationship between the surface mesh units corresponding to each heat exchange path unit; When the cooling tower inlet water enters the packing structure area, the surface grid unit is continuously monitored. When the water film coverage rate changes from less than the preset coverage rate threshold to equal to or greater than the preset coverage rate threshold in the surface grid unit, the corresponding surface grid unit is marked as the water film attachment start position, and the corresponding heat exchange path unit number and timestamp are recorded. According to the preset time sampling interval, the surface grid unit is periodically sampled, and the water film coverage and water film flow direction vector of the surface grid unit are obtained at each sampling time. Between adjacent sampling times, within the same heat exchange path unit, when the water film coverage of adjacent surface grid units is greater than a preset coverage threshold and the angle between the corresponding water film flow direction vectors is less than a preset angle threshold, the adjacent surface grid units are connected sequentially according to the water film flow direction to form a water film flow trajectory segment. The water film flow trajectory segments formed at continuous sampling times are spliced ​​together in chronological order, and the connection relationship across heat exchange path units is established by combining the heat exchange path unit number to form the flow trajectory of the water film along the heat exchange path unit. At the same sampling time, when multiple water film flow trajectory segments converge in the same surface grid cell and the water film coverage is greater than a preset multiple of the average coverage of adjacent surface grid cells, the corresponding surface grid cell is marked as the water film confluence location, and the corresponding heat exchange path cell number and timestamp are recorded. Between adjacent sampling times, when the water film coverage of the surface grid unit changes from greater than the preset coverage threshold to less than the preset fracture threshold, the corresponding surface grid unit is marked as the water film fracture location, and the corresponding heat exchange path unit number and timestamp are recorded. By integrating the water film attachment start position, the water film flow trajectory along the heat exchange path unit, the water film confluence position, and the water film break position in chronological order, a water film attachment trajectory sequence is generated. In the specific implementation process, based on the heat exchange path set of the packing material, the flow state of circulating water on the surface of the packing material is continuously collected. The sampling time interval is set to 0.5s. All surface grid units are periodically scanned to obtain the water film coverage and water film flow direction vector of each surface grid unit. When the water film coverage of a surface grid unit changes from less than 0.7 to equal to or greater than 0.7, the corresponding surface grid unit is marked as the water film attachment start position, and the corresponding heat exchange path unit number and timestamp are recorded. Between adjacent sampling times, within the same heat exchange path unit, the angle between the water film flow direction vectors of adjacent surface grid units is calculated. When the angle is less than 20° and the water film coverage of both surface grid units is equal to the specified value, the flow direction vector is considered to be the flow direction vector. When the water film coverage rate is greater than 0.7, adjacent surface grid cells are sequentially connected according to the water film flow direction to form water film flow trajectory segments. These segments are then spliced ​​together chronologically, and cross-cell connections are established based on the heat exchange path cell number to form the water film flow trajectory along the heat exchange path cell. When multiple water film flow trajectory segments converge in the same surface grid cell and the water film coverage rate of that surface grid cell is greater than 1.5 times the average coverage rate of adjacent surface grid cells, that surface grid cell is marked as the water film confluence location. Furthermore, the breakage threshold is set to 0.3. When the water film coverage rate of a surface grid cell changes from greater than 0.7 to less than 0.3, the corresponding surface grid cell is marked as the water film breakage location. Finally, the water film attachment start position, water film flow trajectory, water film confluence location, and water film breakage location are integrated chronologically to generate a water film attachment trajectory sequence.

[0027] In this embodiment, the tension topology construction module is specifically as follows: Read the water film attachment trajectory sequence and integrate it in chronological order: the water film attachment start position, the water film flow trajectory along the heat exchange path unit, the water film confluence position, and the water film break position. Using the initial position of water film adhesion as the starting point, the corresponding surface grid units are sequentially expanded based on the flow trajectory of the water film along the heat transfer path unit to construct the temporal distribution chain of the surface grid units within each heat transfer path unit. Within the same heat exchange path unit, the changes in water film coverage and the changes in the angle between water film flow direction vectors of each surface grid unit in the time-series distribution chain are calculated between adjacent sampling times. When the change in water film coverage of a surface grid cell between several consecutive sampling times is less than or equal to a preset stability threshold, and the change in the angle between the water film flow direction vectors is less than or equal to a preset change threshold, and the surface grid cell does not belong to the water film confluence location or the water film breakage location, the corresponding surface grid cell is marked as a water film stable region. When a surface grid cell is located on the flow trajectory of the water film along the heat transfer path cell, and the increase in water film coverage between several consecutive sampling times is greater than a preset stability threshold, or the surface grid cell is adjacent to the water film confluence position and the change in the angle between the water film flow direction vectors is greater than a preset change threshold, the corresponding surface grid cell is marked as a water film stretching region. When a surface grid cell is located at the end of the flow trajectory of the water film along the heat transfer path cell, and the reduction in water film coverage between several consecutive sampling times is greater than a preset stability threshold, or when the surface grid cell is adjacent to the water film breakage position and the change in the angle between the water film flow direction vectors is greater than a preset change threshold, the corresponding surface grid cell is marked as a water film contraction region. Within the same heat exchange path unit, spatially adjacent and identically labeled surface mesh units are aggregated to form tension nodes, and the heat exchange path unit number, three-dimensional coordinates, and label type corresponding to each tension node are recorded. Based on the time sequence from the initial position of water film attachment to the water film confluence position, and from the water film confluence position to the water film breakage position, a directed connection relationship is established between tension nodes to generate a water film tension topology map. In the specific implementation process, further spatiotemporal analysis and processing of the packing surface are performed based on the water film adhesion trajectory sequence. The water film adhesion initiation position, the water film flow trajectory along the heat exchange path unit, the water film confluence position, and the water film breakage position are indexed in chronological order to form time series data, with the sampling time interval maintained at 0.5s. A time series distribution chain is constructed for the surface grid units within the same heat exchange path unit, and the change in water film coverage between adjacent sampling times is calculated. The stability threshold is set to 0.05, and the change threshold is set to 10°. When the change in water film coverage within three consecutive sampling times is less than or equal to 0.05 and the change in the angle between the water film flow direction vectors is less than or equal to 10°, and the surface... When a surface grid cell does not belong to a water film confluence location or a water film breakage location, the corresponding surface grid cell is identified as a stable water film region. When the increase in coverage is greater than 0.05 within three consecutive sampling times, or when the surface grid cell is adjacent to a water film confluence location and the change in the angle between the water film flow direction vectors is greater than 10°, it is marked as a water film stretching region. When the decrease in coverage is greater than 0.05 within three consecutive sampling times, or when the surface grid cell is adjacent to a water film breakage location and the change in the angle between the water film flow direction vectors is greater than 10°, it is marked as a water film contraction region. Subsequently, adjacent surface grid cells with consistent markings are aggregated, and directed connections between tension nodes are established in chronological order to generate a water film tension topology map.

[0028] In this embodiment, the environment-driven thermal potential construction module specifically comprises: External environmental parameters, including ambient temperature, ambient humidity, and ambient wind speed, are collected on the air inlet side of the cooling tower. In the three-dimensional coordinate model, the plane where the air inlet of the cooling tower is located is selected as the air inlet section, and it is divided into multiple environmental areas according to the preset grid size. Each environmental area is numbered and its corresponding spatial coordinates are recorded. The ambient temperature, ambient humidity, and ambient wind speed are assigned to the corresponding environmental area. When there are multiple sampling values ​​in the same environmental area, the arithmetic mean of each sampling value is used to obtain the ambient area temperature, ambient area humidity, and ambient area wind speed. The ambient temperature and humidity are converted to obtain the wet-bulb temperature of the ambient area. Specifically, the ambient humidity is converted into a relative humidity ratio between 0 and 1, and the natural logarithm of the relative humidity ratio is performed. The ambient temperature is multiplied by a first preset constant of 17.27, and the result of the multiplication is divided by the sum of the ambient temperature and a second preset constant of 237.7 to obtain the temperature function value. The result of the natural logarithm operation is added to the temperature function value to obtain an intermediate parameter value. The intermediate parameter value is then multiplied by the second preset constant of 237.7, and the result of the multiplication is divided by the difference between the first preset constant of 17.27 and the intermediate parameter value to obtain the wet-bulb temperature of the ambient area. The temperature difference of the environmental area is obtained by subtracting the wet-bulb temperature of the environmental area from the temperature of the environmental area, and the air volume flow rate is obtained by multiplying the wind speed of the environmental area by the area of ​​the environmental area. The heat dissipation capacity of the environmental area is obtained by multiplying the temperature difference of the environmental area by the air volume flow rate. Arrange the heat dissipation capacity values ​​of each environmental region according to spatial location to construct an environmental-driven thermal potential map. In the specific implementation process, multiple environmental parameter collection points are set on the air inlet side of the cooling tower to collect ambient temperature, ambient humidity, and ambient wind speed, with a sampling time interval of 0.5s. In the three-dimensional coordinate model, the plane where the air inlet is located is selected as the air inlet section, and it is divided into multiple environmental regions according to a grid size of 5mm×5mm. Each environmental region is given a unique number and its corresponding three-dimensional coordinates are recorded. The spatial resolution of each environmental region is consistent with the surface grid unit in the packing heat exchange path set, and a corresponding relationship is established, so that the environmental driven thermal potential map and the packing heat exchange path set are directly matched in space without the need for interpolation or resampling, thereby reducing data conversion errors. The ambient temperature, ambient humidity, and ambient wind speed are arithmetically averaged within the same environmental region to obtain the environmental region temperature, environmental region humidity, and environmental region wind speed. By unifying the spatial scale and time sampling interval, the environmental parameters and the water film attachment trajectory sequence are synchronously aligned, improving the accuracy and stability of tension-thermal potential coupling analysis.

[0029] In this embodiment, the tension-thermal potential coupling module is specifically as follows: Based on the tension nodes formed by the aggregation of multiple surface mesh units in the water film tension topology diagram, the heat transfer path unit number, three-dimensional coordinates and label type corresponding to each tension node are read, and the directed connection relationship between tension nodes is obtained according to the time sequence from the water film attachment start position to the water film confluence position, and from the water film confluence position to the water film breakage position. Read the spatial coordinates and heat dissipation capacity values ​​of each environmental region from the environmental-driven thermal potential map; For each tension node, all surface mesh elements in the tension node are extracted one by one, and the three-dimensional coordinates of each surface mesh element are projected along the air inlet direction onto the plane where the air inlet section is located to obtain the projected coordinates; Based on the position of the projected coordinates in the air inlet section, determine the corresponding environmental area and read the heat dissipation capacity value of the environmental area; The heat dissipation capacity values ​​corresponding to all surface mesh elements within the tension node are accumulated and averaged to obtain the equivalent heat dissipation capacity value of the corresponding tension node. The equivalent heat dissipation capacity value is associated with the label type of the corresponding tension node, and the tension-thermal potential coupling structure diagram is generated by combining the directed connection relationship between the tension nodes. In this invention, the tension nodes are spatially mapped based on the water film tension topology map. Each tension node consists of multiple surface mesh units, each 5mm × 5mm in size, corresponding to the spatial resolution of the environmental region in the environmental driven thermal potential map. The three-dimensional coordinates of all surface mesh units within the tension node are orthogonally projected along the air inlet direction, which is consistent with the Y-axis in the three-dimensional coordinate model. The Y-axis component of the three-dimensional coordinates is eliminated, leaving only the X and Z axes as the projection coordinates. The corresponding environmental region is matched based on the position of the projection coordinates on the air inlet section, and the heat dissipation capacity value of that environmental region is read. The heat dissipation capacity values ​​corresponding to all surface mesh units within the same tension node are accumulated and divided by the number of surface mesh units to obtain the equivalent heat dissipation capacity value. This achieves a precise correspondence between the three-dimensional tension structure and the two-dimensional environmental thermal potential, avoiding spatial interpolation errors and improving the calculation accuracy and stability of the coupled structure diagram.

[0030] In this embodiment, the gas-liquid coupling state modeling module specifically comprises: Based on the tension-thermal potential coupling structure diagram, the heat transfer path unit number, equivalent heat dissipation capacity value and marking type corresponding to each tension node are read; The tension nodes of each heat exchange path unit are classified according to the marking type of the tension nodes. The marking types include water film stable region, water film stretching region and water film contraction region. At each sampling time, the equivalent heat dissipation capacity values ​​of the tension nodes corresponding to the stable region of the water film in each heat exchange path unit are accumulated to obtain the stable heat dissipation term; the equivalent heat dissipation capacity values ​​of the tension nodes corresponding to the stretched region of the water film are accumulated to obtain the stretched heat dissipation term; and the equivalent heat dissipation capacity values ​​of the tension nodes corresponding to the contracted region of the water film are accumulated to obtain the contracted heat dissipation term. Multiply the stretching heat dissipation term by a set enhancement coefficient, multiply the contraction heat dissipation term by a set weakening coefficient, and add the enhanced stretching heat dissipation term, the weakened contraction heat dissipation term, and the stable heat dissipation term to obtain the gas-liquid coupling state characteristic value of each heat exchange path unit at the current sampling time. At each sampling moment, the gas-liquid coupling state feature values ​​corresponding to each heat exchange path unit are spliced ​​together according to the heat exchange path unit number order to obtain the path state vector at the current sampling moment. The path state vectors obtained at different sampling times are arranged in chronological order to generate a gas-liquid coupling state sequence; In the specific implementation process, by distinguishing the labeling types of tension nodes and calculating the equivalent heat dissipation capacity contributions corresponding to the stable water film region, the stretched water film region, and the contracted water film region respectively, the influence of different water film morphologies on the heat transfer process can be quantitatively expressed, thereby avoiding the information loss caused by the uniform processing of gas-liquid states in traditional methods. By performing item-by-item accumulation and combination calculations on multiple types of tension nodes in the same heat transfer path unit at each sampling time, with the enhancement coefficient set to 1.2 and the weakening coefficient set to 0.6, the enhancement and weakening effects of water film expansion and contraction on heat transfer capacity can be distinguished and expressed, thereby improving the physical consistency of gas-liquid state characterization. By splicing the gas-liquid coupling state feature values ​​of each heat transfer path unit according to the numbering order of the heat transfer path unit, a path state vector is formed, so that spatial distribution information is preserved in the vector structure, avoiding the loss of spatial structure information. Furthermore, by arranging the path state vectors at different sampling times in a time sequence, the evolution process of the gas-liquid coupling state in the time dimension is continuously described, thereby providing a time-series input for the model to predict the temperature change trend of the cooling tower, improving the stability and response accuracy of the prediction results.

[0031] In this embodiment, the temperature trend prediction module specifically includes: The tension-thermal potential coupling structure diagram, gas-liquid coupling state sequence and historical outlet water temperature data are input into the state encoding unit. The equivalent heat dissipation capacity value, label type and heat exchange path unit number corresponding to each tension node are extracted according to the tension node order. The label type is numerically encoded to form a node feature vector. Align the path state vector in the gas-liquid coupling state sequence with the historical effluent temperature data according to the sampling time, and connect the path state vector at the same sampling time with the corresponding effluent temperature to form a time feature vector; match the node feature vector of each tension node with the time feature vector at the corresponding sampling time and connect them to generate the input feature sequence. The input feature sequence is fed into the dynamic feature mapping unit. Based on the directed connection relationship between tension nodes, each tension node and its adjacent tension nodes with directed connection relationship are taken as neighborhood nodes. A three-layer graph convolution structure is used for feature propagation. The adjacency range of each graph convolution layer is set to the first-order neighborhood, the convolution kernel size is set to 3×3, and the number of output channels is set to 32, 64 and 64 respectively. Linear transformation and ReLU nonlinear activation processing are performed on the output of each layer to generate dynamic features. The dynamic features are input into the spectral evolution analysis unit, and the dynamic spectral layer evolution mechanism is executed to generate spectral coupling features; The spectral coupling features are input into the Koopman state evolution unit, arranged in the order of sampling time to form a state feature sequence, and the spectral coupling features corresponding to each sampling time are used as the state vector. The predicted state vector for the next sampling time is obtained by multiplying the state vector at the current sampling time with the state evolution operator. The predicted state vector is then appended to the end of the state vector sequence in chronological order. At the same time, the state vector corresponding to the earliest sampling time is removed, forming an updated state vector sequence. Based on the updated state vector sequence, the multiplication operation between the state vector and the state evolution operator is repeatedly performed to realize the step-by-step recursion of the state vector sequence and generate the state feature sequence at multiple future sampling times; The state feature sequence at multiple future sampling times is input into the prediction output unit, which consists of a fully connected layer and an output mapping layer. The fully connected layer performs feature mapping on the state feature sequence at multiple future sampling times, and the number of neurons in the fully connected layer is set to 64. The output mapping layer performs a linear transformation on the output of the fully connected layer to obtain the predicted value of the effluent temperature at the corresponding sampling time. The predicted values ​​of the effluent temperature at multiple sampling times are arranged in chronological order to output the future temperature change trend. In this invention, the improved deep Koopman network retains the basic idea of ​​mapping nonlinear systems to a high-dimensional linear space and performing linear state evolution, as is the fundamental principle of existing deep Koopman networks. It models complex dynamic processes through state encoding, feature mapping, and linear evolution. Building upon this, the invention improves the input structure and feature representation methods to address the spatial distribution and temporal evolution characteristics of the gas-liquid coupling process in cooling towers. Specifically, it uses the tension-thermal potential coupling structure diagram as input, enabling the model to simultaneously characterize the topological relationships between tension nodes during state mapping, thereby enhancing its ability to express spatial coupling effects. Simultaneously, it unifies the encoding of the gas-liquid coupling state sequence with historical outlet water temperature data, allowing the fusion of multi-source information in the temporal dimension within the same state space, improving the ability to interpret dynamic trends. The model achieves higher characterization accuracy. Furthermore, by setting up a spectral evolution analysis unit and introducing a dynamic spectral hierarchical evolution mechanism, it processes features with different frequency variations hierarchically and establishes inter-layer coupling relationships. This allows the model to simultaneously capture slow trends, transitional processes, and rapid perturbation behaviors, compensating for the insufficient ability of traditional deep Koopman networks to distinguish multi-scale dynamic features. Further, by setting up a Koopman state evolution unit, the spectral coupling features are transformed into state vectors with a unified structure and recursively updated under the same evolution operator. This ensures consistency in state transition relationships between different sampling times, avoiding the problem of gradual error accumulation and amplification during multi-step prediction. Simultaneously, it enhances the model's ability to stably characterize long-term sequence evolution processes, ensuring smoothness and coherence of prediction results across continuous time intervals. These improvements enable the state evolution process to not only possess linear predictability but also the ability to hierarchically express complex coupled dynamics, significantly improving the stability and accuracy of future temperature change trend predictions, reducing dependence on single-time-point data, and enhancing the model's adaptability and generalization performance under different operating conditions.

[0032] In this embodiment, the dynamic spectrum hierarchical evolution mechanism is specifically as follows: The difference between the dynamic features at adjacent sampling times is calculated. The dynamic features at the next sampling time are subtracted from the dynamic features at the previous sampling time. The absolute value of the difference result is processed. The absolute value results of each component in the same dynamic feature are accumulated and divided by the number of components to obtain the feature change of the corresponding dynamic feature. When the feature change is less than or equal to the set low-frequency threshold, the corresponding dynamic feature is classified as a low-frequency layer; when the feature change is greater than the set low-frequency threshold and less than the set high-frequency threshold, the corresponding dynamic feature is classified as a mid-frequency layer; when the feature change is greater than or equal to the set high-frequency threshold, the corresponding dynamic feature is classified as a high-frequency layer. The dynamic features of the low-frequency layer are accumulated time-by-time in chronological order, and the accumulated result is divided by the number of sampling times involved in the accumulation to generate low-frequency stationary features; The difference vector between adjacent sampling times is calculated sequentially for the dynamic features of the mid-frequency layer, where each difference vector is the dynamic feature of the next sampling time minus the dynamic feature of the previous sampling time. The sampling times are numbered in chronological order. The minimum number is subtracted from the number of each sampling time, and then the result is divided by the difference between the maximum and minimum numbers to obtain the normalized time weight. Each difference vector is multiplied by its corresponding normalized time weight and then summed. The summation result is divided by the sum of all normalized time weights to obtain the mid-frequency variation characteristics. The high-frequency layer dynamic features are processed in chronological order. First, the high-frequency layer dynamic features at all sampling times are accumulated component by component and divided by the number of sampling times to obtain the high-frequency layer average features. The deviation features are obtained by subtracting the average features of the high-frequency layer from the dynamic features of the high-frequency layer at each sampling time. By performing component-by-component absolute value processing on the deviation characteristics, the high-frequency amplitude characteristics corresponding to each sampling time are obtained; The high-frequency amplitude features corresponding to all sampling times are accumulated component by component in chronological order, and the accumulated result is divided by the number of sampling times to generate high-frequency disturbance features. The low-frequency stationary characteristics and the mid-frequency variation characteristics are added component by component to generate the mid-frequency coupling characteristics; The mid-frequency variation characteristics and high-frequency disturbance characteristics are added component by component to generate high-frequency coupling characteristics; The low-frequency stable features, mid-frequency coupling features, and high-frequency coupling features are spliced ​​together in order of frequency hierarchy to generate spectral coupling features; In the specific implementation process, the dynamic characteristics of adjacent sampling times are calculated in the time dimension. The absolute value of the difference results is processed and the average is accumulated by components to obtain the characteristic change. The low-frequency threshold is set to 0.08 and the high-frequency threshold is set to 0.25 to achieve frequency stratification of dynamic characteristics. For the low-frequency dynamic characteristics, the components are accumulated and averaged to obtain the low-frequency stationary characteristics, which are used to characterize the slow change trend in the heat exchange process. For the mid-frequency dynamic characteristics, the difference vector between adjacent sampling times is calculated, and the time weight is calculated. Then, the difference vector is multiplied by the corresponding time weight and accumulated to obtain the mid-frequency change characteristics, which are used to characterize the medium-rate change process. For the high-frequency dynamic characteristics, the components of all sampling times are averaged to obtain the high-frequency average characteristics. Then, the deviation characteristics between each sampling time and the high-frequency average characteristics are calculated. The absolute value of the deviation characteristics is processed by components to obtain the high-frequency amplitude characteristics. The high-frequency disturbance characteristics are further generated to characterize the instantaneous fluctuation intensity. By employing the aforementioned layered processing method, dynamic characteristics at different time scales can be independently characterized. Cross-layer information fusion is achieved through mid-frequency coupling and high-frequency coupling, enabling the model to simultaneously capture stable trends, gradual processes, and rapid disturbance behaviors. This significantly improves the modeling accuracy of complex gas-liquid coupled heat transfer processes and enhances the stability and anti-interference capability of temperature change trend prediction.

[0033] In this embodiment, the operating parameter adjustment module specifically comprises: Read future temperature change trends and obtain predicted effluent temperatures at multiple sampling times; The predicted effluent temperatures at multiple sampling times are summed and divided by the number of sampling times to obtain the average predicted temperature. The set target effluent temperature is then subtracted from the average predicted temperature to obtain the temperature deviation value. The operating parameters of the cooling tower are adjusted based on the temperature deviation value. These operating parameters include circulating water flow rate and air flow rate. When the temperature deviation exceeds the set positive threshold, increase the circulating water flow rate and air flow rate; When the temperature deviation is less than the set negative threshold, reduce the circulating water flow rate and air flow rate; The cooling tower is operated and controlled based on the adjusted circulating water flow and air flow to achieve adaptive adjustment of the outlet water temperature. In the specific implementation process, the target outlet water temperature is set to 30℃, the positive threshold is set to 1.5℃, and the negative threshold is set to -1.5℃. When the average predicted temperature is 1.5℃ higher than the target outlet water temperature, it is determined to be an insufficient cooling state, and the circulating water flow and air flow are increased. When the average predicted temperature is 1.5℃ lower than the target outlet water temperature, it is determined to be an over-cooling state, and the circulating water flow and air flow are reduced. By dividing the above threshold range, the cooling intensity can be stably adjusted, avoiding frequent adjustments that cause system fluctuations and improving the stability and response accuracy of the temperature control process.

[0034] Example 1: To verify the feasibility of this invention in practice, it was applied to a large mechanical ventilation cooling tower in an industrial circulating cooling system. This cooling tower operates under conditions of significant load fluctuations and frequent changes in the external environment. In traditional operation modes, the cooling tower mainly relies on manual experience or a simple feedback control method based on outlet water temperature to adjust the circulating water flow and air flow. In actual operation, this results in problems such as large fluctuations in outlet water temperature, slow response, and high energy consumption. Especially when the ambient temperature and humidity change rapidly, the system struggles to adapt in time, leading to a decrease in cooling efficiency.

[0035] In this embodiment, the cooling tower packing structure is first modeled in three dimensions using a packing heat exchange path construction module to establish a set of packing heat exchange paths and form high-precision surface mesh units on the packing surface. Then, the flow state of circulating water on the packing surface is continuously collected using a water film trajectory acquisition module to generate a water film attachment trajectory sequence, so that the water film attachment, confluence, and breakage processes are completely recorded. On this basis, the water film stable region, water film stretching region, and water film contraction region in each heat exchange path unit are identified using a tension topology construction module, and a water film tension topology map is formed, so that the spatial distribution state of the water film is expressed in a structured way.

[0036] Furthermore, the ambient temperature, relative humidity, and wind speed are collected through the environment-driven thermal potential construction module. The cooling tower inlet area is divided into environmental regions, and the heat dissipation capacity of each environmental region is calculated to construct an environment-driven thermal potential map. Subsequently, in the tension-thermal potential coupling module, the water film tension topology map is spatially mapped to the environment-driven thermal potential map, and the heat dissipation capacity of the environmental region is mapped to the corresponding heat exchange path unit to generate a tension-thermal potential coupling structure diagram, thereby realizing a unified modeling of the water film state and the environmental heat exchange capacity.

[0037] In the gas-liquid coupling state modeling module, the gas-liquid coupling state characteristic values ​​of each heat exchange path unit at each sampling time are calculated, and a gas-liquid coupling state sequence is constructed to continuously describe the dynamic changes of the heat exchange process. In the temperature trend prediction module, the tension-thermal potential coupling structure diagram, the gas-liquid coupling state sequence, and historical effluent temperature data are input into an improved deep Koopman network. The network is processed by a state encoding unit, a dynamic feature mapping unit, a spectral evolution analysis unit, a Koopman state evolution unit, and a prediction output unit to output the future temperature change trend. Finally, the operation parameter adjustment module adaptively adjusts the circulating water flow rate and air flow rate according to the future temperature change trend to achieve precise control of the effluent temperature.

[0038] To verify the actual effect of the present invention, the operating data of the traditional control method based on simple feedback of outlet water temperature and the method of the present invention were compared under the same working conditions. The results are shown in Table 1 below.

[0039] Water outlet temperature range (°C) 28.5~33.2 29.6~30.8 Average outlet water temperature (°C) 30.9 30.1 Temperature standard deviation (°C) 1.42 0.46 Maximum temperature fluctuation (°C) 4.7 1.2 System response time (s) 180 75 Average circulating water flow rate (m³ / h) 520 495 Average airflow rate (m³ / h) <![CDATA[1.8×10 5 ]]> <![CDATA[1.65×10 5 ]]> Energy consumption per unit time (kWh) 215 188 As can be seen from the data in Table 1, the method of this invention outperforms the traditional control method in several key operating indicators. Regarding outlet water temperature control, the traditional control method has a temperature range of 28.5℃ to 33.2℃, with a large fluctuation range, while the method of this invention stabilizes the temperature between 29.6℃ and 30.8℃, significantly reducing the temperature fluctuation range; the corresponding average outlet water temperature decreases from 30.9℃ to 30.1℃, closer to the set target outlet water temperature of 30℃; simultaneously, the temperature standard deviation decreases from 1.42℃ to 0.46℃, and the maximum temperature fluctuation decreases from 4.7℃ to 1.2℃, indicating a significant improvement in temperature control stability. In terms of dynamic response, the system response time is shortened from 180s to 75s, indicating that the method of this invention can adapt to changes in operating conditions more quickly. Regarding operational energy efficiency, the circulating water flow rate decreases from 520m³ / h to 495m³ / h, and the air flow rate decreases from 1.8×10⁻⁶ to 1.8×10⁻⁶. 5 m³ / h decreased to 1.65 × 10 5 With a capacity of m³ / h, energy consumption per unit time decreased from 215kWh to 188kWh, achieving energy reduction while maintaining cooling effect, demonstrating superior energy-saving performance.

[0040] In this embodiment, by constructing a tension-thermal potential coupled structure diagram and a gas-liquid coupled state sequence, a unified model of the water film flow state inside the cooling tower and the heat dissipation capacity of the environment is realized, so that the spatial distribution characteristics and temporal evolution law of the heat exchange process can be characterized simultaneously. Combined with an improved deep Koopman network, a dynamic spectrum hierarchical evolution mechanism is introduced to express multi-scale dynamic characteristics in a hierarchical manner and realize stable recursive prediction of the state, so that the temperature change trend has higher predictability and continuity, thereby effectively improving the stability, response capability and overall operating efficiency of the cooling tower temperature control process.

[0041] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning based cooling tower adaptive temperature control system, characterized in that, include: The packing heat exchange path construction module is used to perform three-dimensional modeling of the cooling tower packing structure, generate heat exchange path units, and construct a packing heat exchange path set. The water film trajectory acquisition module is used to continuously acquire the flow state of circulating water on the surface of the packing material based on the heat exchange path set of the packing material, and generate a water film adhesion trajectory sequence. The tension topology construction module is used to spatially divide the packing surface according to the water film attachment trajectory sequence, identify the water film stable region, water film stretching region and water film contraction region in each heat exchange path unit, and generate a water film tension topology map. The environment-driven thermal potential construction module is used to collect external environmental parameters of the cooling tower, divide the environmental areas and calculate the heat dissipation capacity value of each environmental area, and construct an environment-driven thermal potential map. The tension-thermal potential coupling module is used to spatially map the water film tension topology map with the environmental driven thermal potential map, and to map the heat dissipation capacity of the environmental area to the corresponding heat exchange path unit, generating a tension-thermal potential coupling structure map. The gas-liquid coupling state modeling module is used to calculate the gas-liquid coupling state characteristic values ​​of each heat transfer path unit at each sampling time based on the tension-thermal potential coupling structure diagram, and construct the gas-liquid coupling state sequence. The temperature trend prediction module is used to input the tension-thermal potential coupling structure diagram, gas-liquid coupling state sequence and historical effluent temperature data into the improved deep Koopman network, and process them sequentially through the state encoding unit, dynamic feature mapping unit, spectral evolution analysis unit, Koopman state evolution unit and prediction output unit. The spectral evolution analysis unit introduces a dynamic spectral layer evolution mechanism to output the future temperature change trend. The operating parameter adjustment module is used to adjust the operating parameters of the cooling tower according to future temperature change trends, so as to achieve adaptive control of the cooling tower outlet water temperature.

2. The cooling tower adaptive temperature control system based on deep learning according to claim 1, characterized in that, The packing heat exchange path construction module is specifically as follows: Obtain the structural parameters of the cooling tower packing structure, including the packing layer thickness, layer spacing, and layer tilt angle, and establish a three-dimensional coordinate model of the packing structure based on the structural parameters, and establish a three-dimensional rectangular coordinate system with the bottom of the packing as the coordinate origin; The surface of the filler sheet in the three-dimensional coordinate model is discretized and divided into multiple surface mesh units according to a preset spatial resolution. Each surface mesh unit corresponds to a unique three-dimensional coordinate. In each surface grid cell, a gravity direction vector and a sheet tilt angle direction vector are established, and the vectors are superimposed to obtain the water film flow direction vector; Between adjacent surface grid cells, when the angle between two water film flow direction vectors is less than a preset angle threshold, the adjacent surface grid cells are connected sequentially along the water film flow direction to form a water film adhesion path. In the three-dimensional coordinate model, an airflow direction vector field along the air inlet direction is established, and the channels between the packing sheets are divided into multiple volumetric grid units. Between adjacent volume grid cells, when the angle between the airflow direction vectors is less than a preset angle threshold, the adjacent volume grid cells are connected sequentially along the airflow direction to form an airflow path. In the surface grid cells where the water film adhesion path and the air flow path intersect, surface grid cells with a water film coverage rate greater than a preset coverage rate threshold and an air flow velocity greater than a preset flow velocity threshold are selected as water-air contact areas, and the corresponding three-dimensional coordinates are recorded. Centered on the water-air contact area, adjacent surface grid units that satisfy the condition that the angle between the water film flow direction vectors is less than a preset angle threshold and that the air channels are connected are aggregated to form a heat exchange path unit. Each heat exchange path unit encloses at least one water-air contact area. Each heat exchange path unit is numbered and its corresponding three-dimensional coordinates are recorded to generate a set of packing heat exchange paths.

3. The deep learning-based adaptive temperature control system for cooling towers according to claim 1, characterized in that, The water film trajectory acquisition module is specifically as follows: Read the unit number and corresponding three-dimensional coordinates of each heat exchange path in the packing heat exchange path set, and establish an index relationship between the surface mesh units corresponding to each heat exchange path unit; When the cooling tower inlet water enters the packing structure area, the surface grid unit is continuously monitored. When the water film coverage rate changes from less than the preset coverage rate threshold to equal to or greater than the preset coverage rate threshold in the surface grid unit, the corresponding surface grid unit is marked as the water film attachment start position, and the corresponding heat exchange path unit number and timestamp are recorded. According to the preset time sampling interval, the surface grid unit is periodically sampled, and the water film coverage and water film flow direction vector of the surface grid unit are obtained at each sampling time. Between adjacent sampling times, within the same heat exchange path unit, when the water film coverage of adjacent surface grid units is greater than a preset coverage threshold and the angle between the corresponding water film flow direction vectors is less than a preset angle threshold, the adjacent surface grid units are connected sequentially according to the water film flow direction to form a water film flow trajectory segment. The water film flow trajectory segments formed at continuous sampling times are spliced ​​together in chronological order, and the connection relationship across heat exchange path units is established by combining the heat exchange path unit number to form the flow trajectory of the water film along the heat exchange path unit. At the same sampling time, when multiple water film flow trajectory segments converge in the same surface grid cell and the water film coverage is greater than a preset multiple of the average coverage of adjacent surface grid cells, the corresponding surface grid cell is marked as the water film confluence location, and the corresponding heat exchange path cell number and timestamp are recorded. Between adjacent sampling times, when the water film coverage of the surface grid unit changes from greater than the preset coverage threshold to less than the preset fracture threshold, the corresponding surface grid unit is marked as the water film fracture location, and the corresponding heat exchange path unit number and timestamp are recorded. By integrating the water film attachment start position, the water film flow trajectory along the heat exchange path unit, the water film confluence position, and the water film break position in chronological order, a water film attachment trajectory sequence is generated.

4. The deep learning-based adaptive temperature control system for cooling towers according to claim 1, characterized in that, The tension topology construction module is specifically as follows: Read the water film attachment trajectory sequence and integrate it in chronological order: the water film attachment start position, the water film flow trajectory along the heat exchange path unit, the water film confluence position, and the water film break position. Using the initial position of water film adhesion as the starting point, the corresponding surface grid units are sequentially expanded based on the flow trajectory of the water film along the heat transfer path unit to construct the temporal distribution chain of the surface grid units within each heat transfer path unit. Within the same heat exchange path unit, the changes in water film coverage and the changes in the angle between water film flow direction vectors of each surface grid unit in the time-series distribution chain are calculated between adjacent sampling times. When the change in water film coverage of a surface grid cell between several consecutive sampling times is less than or equal to a preset stability threshold, and the change in the angle between the water film flow direction vectors is less than or equal to a preset change threshold, and the surface grid cell does not belong to the water film confluence location or the water film breakage location, the corresponding surface grid cell is marked as a water film stable region. When a surface grid cell is located on the flow trajectory of the water film along the heat transfer path cell, and the increase in water film coverage between several consecutive sampling times is greater than a preset stability threshold, or the surface grid cell is adjacent to the water film confluence position and the change in the angle between the water film flow direction vectors is greater than a preset change threshold, the corresponding surface grid cell is marked as a water film stretching region. When a surface grid cell is located at the end of the flow trajectory of the water film along the heat transfer path cell, and the reduction in water film coverage between several consecutive sampling times is greater than a preset stability threshold, or when the surface grid cell is adjacent to the water film breakage position and the change in the angle between the water film flow direction vectors is greater than a preset change threshold, the corresponding surface grid cell is marked as a water film contraction region. Within the same heat exchange path unit, spatially adjacent and identically labeled surface mesh units are aggregated to form tension nodes, and the heat exchange path unit number, three-dimensional coordinates, and label type corresponding to each tension node are recorded. Based on the chronological order from the initial position of water film attachment to the water film confluence position, and from the water film confluence position to the water film breakage position, a directed connection relationship is established between tension nodes to generate a water film tension topology map.

5. The deep learning-based adaptive temperature control system for cooling towers according to claim 1, characterized in that, The environmentally driven thermal potential construction module is specifically as follows: External environmental parameters are collected on the air inlet side of the cooling tower, including ambient temperature, ambient humidity, and ambient wind speed. In the three-dimensional coordinate model, the plane where the air inlet of the cooling tower is located is selected as the air inlet section, and it is divided into multiple environmental areas according to the preset grid size. Each environmental area is numbered and its corresponding spatial coordinates are recorded. The ambient temperature, ambient humidity, and ambient wind speed are assigned to the corresponding environmental area. When there are multiple sampling values ​​in the same environmental area, the arithmetic mean of each sampling value is used to obtain the ambient area temperature, ambient area humidity, and ambient area wind speed. The ambient temperature and ambient humidity are converted to obtain the ambient wet-bulb temperature. The temperature difference of the environmental area is obtained by subtracting the wet-bulb temperature of the environmental area from the temperature of the environmental area, and the air volume flow rate is obtained by multiplying the wind speed of the environmental area by the area of ​​the environmental area. The heat dissipation capacity of the environmental area is obtained by multiplying the temperature difference of the environmental area by the air volume flow rate. Arrange the heat dissipation capacity values ​​of each environmental area according to spatial location to construct an environmental-driven thermal potential map.

6. The deep learning-based adaptive temperature control system for cooling towers according to claim 1, characterized in that, The tension-thermal potential coupling module is specifically: Based on the tension nodes formed by the aggregation of multiple surface mesh units in the water film tension topology diagram, the heat transfer path unit number, three-dimensional coordinates and label type corresponding to each tension node are read, and the directed connection relationship between tension nodes is obtained according to the time sequence from the water film attachment start position to the water film confluence position, and from the water film confluence position to the water film breakage position. Read the spatial coordinates and heat dissipation capacity values ​​of each environmental region from the environmental-driven thermal potential map; For each tension node, all surface mesh elements in the tension node are extracted one by one, and the three-dimensional coordinates of each surface mesh element are projected along the air inlet direction onto the plane where the air inlet section is located to obtain the projected coordinates; Based on the position of the projected coordinates in the air inlet section, the corresponding environmental area is determined, and the heat dissipation capacity value of the environmental area is read. The heat dissipation capacity values ​​corresponding to all surface mesh elements within the tension node are accumulated and averaged to obtain the equivalent heat dissipation capacity value of the corresponding tension node. The equivalent heat dissipation capacity value is associated with the label type of the corresponding tension node, and the tension-thermal potential coupling structure diagram is generated by combining the directed connection relationship between the tension nodes.

7. The cooling tower adaptive temperature control system based on deep learning according to claim 1, characterized in that, The gas-liquid coupling state modeling module is specifically as follows: Based on the tension-thermal potential coupling structure diagram, the heat transfer path unit number, equivalent heat dissipation capacity value and marking type corresponding to each tension node are read; The tension nodes of each heat exchange path unit are classified according to the marking type of the tension nodes. The marking type includes water film stable region, water film stretching region and water film contraction region. At each sampling time, the equivalent heat dissipation capacity values ​​of the tension nodes corresponding to the stable region of the water film in each heat exchange path unit are accumulated to obtain the stable heat dissipation term; the equivalent heat dissipation capacity values ​​of the tension nodes corresponding to the stretched region of the water film are accumulated to obtain the stretched heat dissipation term; and the equivalent heat dissipation capacity values ​​of the tension nodes corresponding to the contracted region of the water film are accumulated to obtain the contracted heat dissipation term. Multiply the stretching heat dissipation term by a set enhancement coefficient, multiply the contraction heat dissipation term by a set weakening coefficient, and add the enhanced stretching heat dissipation term, the weakened contraction heat dissipation term, and the stable heat dissipation term to obtain the gas-liquid coupling state characteristic value of each heat exchange path unit at the current sampling time. At each sampling moment, the gas-liquid coupling state feature values ​​corresponding to each heat exchange path unit are spliced ​​together according to the heat exchange path unit number order to obtain the path state vector at the current sampling moment. The path state vectors obtained at different sampling times are arranged in chronological order to generate a gas-liquid coupling state sequence.

8. The cooling tower adaptive temperature control system based on deep learning according to claim 1, characterized in that, The temperature trend prediction module is specifically as follows: The tension-thermal potential coupling structure diagram, gas-liquid coupling state sequence and historical outlet water temperature data are input into the state encoding unit. The equivalent heat dissipation capacity value, label type and heat exchange path unit number corresponding to each tension node are extracted according to the tension node order. The label type is numerically encoded to form a node feature vector. Align the path state vector in the gas-liquid coupling state sequence with the historical effluent temperature data according to the sampling time, and connect the path state vector at the same sampling time with the corresponding effluent temperature to form a time feature vector; match the node feature vector of each tension node with the time feature vector at the corresponding sampling time and connect them to generate the input feature sequence. The input feature sequence is fed into the dynamic feature mapping unit. Based on the directed connection relationship between tension nodes, each tension node and its adjacent tension nodes with directed connection relationship are taken as neighborhood nodes. A three-layer graph convolution structure is used for feature propagation. The adjacency range of each graph convolution layer is set to the first-order neighborhood, the convolution kernel size is set to 3×3, and the number of output channels is set to 32, 64 and 64 respectively. Linear transformation and ReLU nonlinear activation processing are performed on the output of each layer to generate dynamic features. The dynamic features are input into the spectral evolution analysis unit, and the dynamic spectral layer evolution mechanism is executed to generate spectral coupling features; The spectral coupling features are input into the Koopman state evolution unit, arranged in the order of sampling time to form a state feature sequence, and the spectral coupling features corresponding to each sampling time are used as the state vector. The predicted state vector for the next sampling time is obtained by multiplying the state vector at the current sampling time with the state evolution operator. The predicted state vector is then appended to the end of the state vector sequence in chronological order. At the same time, the state vector corresponding to the earliest sampling time is removed to form an updated state vector sequence. Based on the updated state vector sequence, the multiplication operation between the state vector and the state evolution operator is repeatedly performed to realize the step-by-step recursion of the state vector sequence and generate the state feature sequence at multiple future sampling times; The state feature sequences of multiple future sampling times are input into the prediction output unit, which consists of a fully connected layer and an output mapping layer. The fully connected layer performs feature mapping on the state feature sequences of multiple future sampling times, and the number of neurons in the fully connected layer is set to 64. The output mapping layer performs a linear transformation on the output of the fully connected layer to obtain the predicted value of the effluent temperature at the corresponding sampling time. The predicted values ​​of the effluent temperature at multiple sampling times are arranged in chronological order to output the future temperature change trend.

9. The cooling tower adaptive temperature control system based on deep learning according to claim 8, wherein, The specific mechanism of the dynamic spectrum hierarchical evolution is as follows: The difference between the dynamic features at adjacent sampling times is calculated. The dynamic features at the next sampling time are subtracted from the dynamic features at the previous sampling time. The absolute value of the difference result is processed. The absolute value results of each component in the same dynamic feature are accumulated and divided by the number of components to obtain the feature change of the corresponding dynamic feature. When the feature change is less than or equal to the set low-frequency threshold, the corresponding dynamic feature is classified as a low-frequency layer; when the feature change is greater than the set low-frequency threshold and less than the set high-frequency threshold, the corresponding dynamic feature is classified as a mid-frequency layer; when the feature change is greater than or equal to the set high-frequency threshold, the corresponding dynamic feature is classified as a high-frequency layer. The dynamic features of the low-frequency layer are accumulated time-by-time in chronological order, and the accumulated result is divided by the number of sampling times involved in the accumulation to generate low-frequency stationary features; The difference vector between adjacent sampling times is calculated sequentially for the dynamic features of the mid-frequency layer, where each difference vector is the dynamic feature of the next sampling time minus the dynamic feature of the previous sampling time. The sampling times are numbered in chronological order. The minimum number is subtracted from the number of each sampling time, and then the result is divided by the difference between the maximum and minimum numbers to obtain the normalized time weight. Each difference vector is multiplied by its corresponding normalized time weight and then summed. The summation result is divided by the sum of all normalized time weights to obtain the mid-frequency variation characteristics. The high-frequency layer dynamic features are processed in chronological order. First, the high-frequency layer dynamic features at all sampling times are accumulated component by component and divided by the number of sampling times to obtain the high-frequency layer average features. The deviation features are obtained by subtracting the average features of the high-frequency layer from the dynamic features of the high-frequency layer at each sampling time. By performing component-by-component absolute value processing on the deviation characteristics, the high-frequency amplitude characteristics corresponding to each sampling time are obtained; The high-frequency amplitude features corresponding to all sampling times are accumulated component by component in chronological order, and the accumulated result is divided by the number of sampling times to generate high-frequency disturbance features. The low-frequency stationary characteristics and the mid-frequency variation characteristics are added component by component to generate the mid-frequency coupling characteristics; The mid-frequency variation characteristics and high-frequency disturbance characteristics are added component by component to generate high-frequency coupling characteristics; The low-frequency stable features, mid-frequency coupling features, and high-frequency coupling features are spliced ​​together in order of frequency hierarchy to generate spectral coupling features.

10. The cooling tower adaptive temperature control system based on deep learning according to claim 1, wherein, The operating parameter adjustment module is specifically as follows: Read future temperature change trends and obtain predicted effluent temperatures at multiple sampling times; The predicted effluent temperatures at multiple sampling times are summed and divided by the number of sampling times to obtain the average predicted temperature. The set target effluent temperature is then subtracted from the average predicted temperature to obtain the temperature deviation value. The operating parameters of the cooling tower are adjusted based on the temperature deviation value. These operating parameters include circulating water flow rate and air flow rate. When the temperature deviation exceeds the set positive threshold, increase the circulating water flow rate and air flow rate; When the temperature deviation is less than the set negative threshold, reduce the circulating water flow rate and air flow rate; The cooling tower is operated and controlled based on the adjusted circulating water flow and air flow to achieve adaptive adjustment of the outlet water temperature.

Images