Integrated intelligent temperature control and heat dissipation system of prepackaged box-type substation

Abstract

This application relates to an integrated intelligent temperature control and heat dissipation system for a prefabricated box-type substation, belonging to the field of power equipment heat dissipation technology. The system includes a base 1 and a housing 2 mounted on the base 1. The housing 2 is divided into a low-pressure chamber, a transformer chamber, and a high-pressure chamber. The transformer chamber is equipped with an integrated cooling unit 4 connected to a closed-loop temperature control system. The system also includes an intelligent control unit 9. The closed-loop temperature control system includes a cooling plate 5, a heat exchange unit 6, a circulating pump 7, an expansion tank 14, a water replenishment device 15, and an optional plate heat exchanger 8. The intelligent control unit 9 is equipped with a load feedforward-temperature feedback dual-mode predictive controller. This controller embeds a lightweight LSTM model. Input variables include the transformer primary current harmonic distortion rate (THD), the active power change rate (dP / dt), the winding temperature rise history curve measured by distributed optical fiber, and the cooling water inlet and outlet temperature difference (ΔT). The output is the target flow increment (ΔQ) of each cooling chamber within the next 60 seconds and the pre-speed command of the heat exchange fan. This application can solve the problem of control response lag caused by the thermal inertia of the cooling medium and improve the thermal response capability of the transformer under sudden short-term overload conditions.

Description

Technical Field

[0001] This application relates to the field of power equipment heat dissipation technology, specifically to an integrated intelligent temperature control and heat dissipation system for a prefabricated box-type substation. Background Technology

[0002] Prefabricated substations, as crucial power distribution equipment in power systems, are widely used in urban power grids, industrial areas, and residential areas. Their operational stability directly impacts power supply reliability. With the continuous growth of power load and the trend towards compact equipment design, transformer heat dissipation has become increasingly prominent. Current technologies commonly employ air-cooling, water-cooling, or combined cooling systems for temperature control in prefabricated substations. Typical solutions include installing cooling plates in the transformer room, using a circulating pump to drive the cooling medium through a closed-loop pipeline, and achieving heat exchange via an external heat exchange unit. Expansion tanks and water replenishment devices are also included to maintain stable system pressure. Some systems further incorporate plate heat exchangers to improve heat exchange efficiency. Intelligent control units typically use temperature sensor feedback signals to perform closed-loop regulation of the circulating pump and heat exchange fans, enabling monitoring and management of the transformer's operating temperature.

[0003] However, under the condition of sudden short-term overload of transformer, the thermal inertia of the cooling medium causes a significant delay in temperature response, which makes it impossible for the control system to increase the cooling power in time, affecting the safe operation of the equipment. Summary of the Invention

[0004] This application provides an integrated intelligent temperature control and heat dissipation system for prefabricated box-type substations, which can solve the technical problem of control lag caused by the thermal inertia of the cooling medium, affecting the safe operation of the transformer.

[0005] To achieve the above objectives, this application provides the following technical solution: This application provides an integrated intelligent temperature control and heat dissipation system for a prefabricated box-type substation, including a base 1 and a box 2 mounted on the base 1. The box 2 is divided into a low-pressure chamber, a transformer chamber, and a high-pressure chamber by a partition 3. An integrated cooling unit 4 is installed in the transformer chamber, and the integrated cooling unit 4 is connected to a closed-loop temperature control system. The system also includes an intelligent control unit 9 for coordinating the operation of each subsystem. The closed-loop temperature control system includes: a cooling plate 5, installed in the transformer chamber and facing the transformer equipment, internally divided into multiple cooling chambers with air inlets between the cooling chambers; and a heat exchange unit 6, located outside the box 2, including a closed cooling tower or a dry heat exchanger, with a built-in variable frequency fan 1. 8; Circulation pump 7, connected to the outlet of cooling plate 5 and the inlet of heat exchange unit 6, drives cooling water circulation; Expansion tank 14 and water replenishment device 15; Plate heat exchanger 8, which can be optionally installed in the cooling water circuit; Intelligent control unit 9 is equipped with a load feedforward-temperature feedback dual-mode predictive controller. The controller embeds a lightweight long short-term memory neural network LSTM model. The input variables include the transformer primary current harmonic distortion rate THD, the active power change rate dP / dt, the winding temperature rise history curve measured by distributed optical fiber, and the cooling water inlet and outlet temperature difference ΔT. The output is the target flow increment ΔQ of each cooling chamber in the next 60 seconds and the pre-speed command of the variable frequency fan 18 of heat exchange unit 6.

[0006] In one optional embodiment, the cooling plate 5 adopts a modular and detachable structure, which is composed of multiple independent cooling units spliced ​​together in the horizontal direction. Each cooling unit corresponds to a cooling chamber. The units are connected to the water circuit through an axial floating sealing joint. The axial floating sealing joint has a built-in disc spring, which allows ±0.5mm axial displacement to compensate for thermal expansion and contraction. A Venturi-diaphragm type differential pressure regulating valve 11 is integrated at the water inlet of each cooling unit. The Venturi-diaphragm type differential pressure regulating valve 11 uses the kinetic energy of water to drive the elastic diaphragm to deform and automatically adjust the throttling area so that the pressure loss difference between any two adjacent cooling units is constant ≤0.8kPa. The cross-section of the internal flow channel of each cooling unit is gradient-contracted according to the golden ratio λ=0.618.

[0007] In one optional embodiment, the air inlet of the cooling plate 5 is provided with a multi-layer protective structure, and the surface of the condenser plate 12 is divided into three functional areas along the airflow direction: the air inlet side 1 / 3 area is coated with a fluorosilane superhydrophobic coating with a contact angle >160° and a roll-off angle <3°; the middle 1 / 3 area is laser micro-textured to form a hydrophilic area with a micro-pit array diameter of 20μm and a depth of 5μm; the air outlet side 1 / 3 area is coated with a graphene-titanium dioxide photocatalytic hydrophobic layer; a capacitive continuous liquid level sensor 13 is connected to the bottom of the condenser plate 12 with a measurement resolution of 0.1mm, and its output signal is linked to a micro-drainage pump to achieve precise drainage at the drop level.

[0008] In one optional embodiment, the cooling medium used in the closed-loop temperature control system is an aqueous ethylene glycol solution. The expansion tank 14 also integrates a paraffin-based phase change cold storage module 16 with a phase change temperature of 5°C. The circulation pump 7 is a magnetically coupled variable frequency pump with an impeller made of carbon fiber reinforced PEEK. A micro-heating wire mesh is embedded within the impeller flow channel, and the power density is 0.8 W / cm². 2 The intelligent control unit 9 is also equipped with a viscosity compensation mode: when the cooling water temperature is detected to be <-10℃ and the real-time flow rate is reduced by >15% compared with the rated value, the micro heating wire is activated to locally heat the 0.5mm liquid layer around the impeller, so that the dynamic viscosity of the area is reduced by ≥40%.

[0009] In one optional embodiment, the structural parameters of the LSTM model satisfy the following: number of hidden layer nodes ≤ 64, training dataset is derived from measured time series data under at least three typical working conditions, including: sudden addition of 50% rated load under no-load conditions, short-term impact overload (duration ≤ 3s), day-night temperature difference cycle (-25℃ to 40℃), model weight file size < 50KB, deployed in PLC edge controller, and single inference latency < 200ms.

[0010] In one optional embodiment, a distributed fiber optic temperature sensor 17 is attached to the surface of the cooling plate 5. The sensor has a spatial resolution of 10 cm and a temperature measurement accuracy of ±0.5℃. The data from the fiber optic sensor is fused with wavelet denoising and Kalman filtering and used as the input source for the "winding temperature rise history curve" in the LSTM model.

[0011] In an optional embodiment, the variable frequency fan 18 of the heat exchange unit 6 is an EC fan, and its speed control strategy includes: basic mode: closed-loop adjustment based on the cooling water return temperature; feedforward enhancement mode: receiving the pre-speed-up command output by the LSTM model and increasing the fan speed before the return water temperature rises; natural cooling intervention mode: when the ambient wet-bulb temperature is lower than the cooling water set temperature and ΔT>3K, the circulation pump 7 is turned off, and only the fan is used for air-water natural convection heat exchange.

[0012] In an optional embodiment, the direct cooling structure of the transformer body further includes: a spiral coil cooling coil 19 is added inside the transformer oil tank, and its oil inlet and outlet are connected to the cooling water circuit of the closed-loop temperature control system through a plate heat exchanger 8. The ratio of the spiral coil pitch to the tube diameter is 3.2:1, and the inner wall of the tube is provided with a nano-level TiO2 photocatalytic coating to inhibit the growth of microorganisms in the oil circuit.

[0013] In one optional embodiment, the human-machine interface of the intelligent control unit 9 is equipped with a three-dimensional thermal cloud map display module, whose data sources include: a distributed optical fiber temperature field on the surface of the cooling plate 5, a transformer winding fiber optic grating temperature sensor array 20, and pressure and flow sensors 10 at the inlet and outlet of each cooling chamber. The three-dimensional thermal cloud map supports backtracking by time axis, automatic hot spot labeling, and early warning of cooling efficiency decay trend.

[0014] In one optional embodiment, when the system operates continuously within an ambient temperature range of -30℃ to 40℃, it meets the following performance indicators: under a sudden short-term overload (120% rated load, lasting 2s), the peak temperature rise of the transformer winding is reduced by ≥25% compared to when LSTM feedforward control is not enabled; the actual flow deviation of each cooling chamber is ≤±3.7%; the air intake efficiency retention rate is ≥92% in a high humidity environment (RH≥90%, 25℃); the winter mode start-up time is ≤5min; the overall energy consumption is reduced by ≥30% compared to the traditional air-cooled system; and it is reduced by ≥22% compared to the original scheme in Appendix 1.

[0015] This application provides an integrated intelligent temperature control and heat dissipation system for a prefabricated box-type substation. This scheme achieves early prediction of transformer temperature rise trends through a load feedforward-temperature feedback dual-mode predictive controller. By using the transformer primary current harmonic distortion rate (THD) and active power change rate (dP / dt) as precursor indicators of temperature rise, a mapping relationship between electrical parameters and thermal response is established, enabling the controller to issue control commands before the actual occurrence of winding temperature rise. Based on the historical winding temperature rise curve measured by distributed optical fiber and the inlet and outlet temperature difference ΔT of cooling water, the LSTM model can accurately predict the heat load change trend of each cooling chamber within the next 60 seconds. This design allows the system to output the target flow increment ΔQ for each cooling chamber, realizing on-demand spatial and temporal allocation of cooling resources and avoiding energy waste caused by ineffective circulation. With the help of the pre-speed-up command of the variable frequency fan 18 of heat exchange unit 6 output by the LSTM model, the system can increase the fan speed before the return water temperature rise trend appears, compensating for the slow response of the water circuit and forming a water-air dual-channel collaborative accelerated heat dissipation mechanism. This solution transforms the traditional passive response mechanism of "adjusting after temperature rises" into an active feedforward mechanism of "pre-adjusting upon load change," effectively solving the control response lag problem caused by the thermal inertia of the cooling medium. Through the design of a lightweight LSTM model (<50KB, <200ms), the real-time operation capability of the algorithm on industrial edge controllers is ensured, enabling intelligent feedforward control to move from theory to engineering application. This system significantly improves the thermal response capability of prefabricated substations under sudden short-term overload conditions, ensuring the safe and stable operation of transformers. Attached Figure Description

[0016] Figure 1 A schematic diagram of the structure of an integrated intelligent temperature control and heat dissipation system for a prefabricated box-type substation provided in this application; Figure 2 A process flow diagram of the closed-loop temperature control system provided in this application; Figure 3 A cross-sectional structural diagram of the cooling plate 5 provided in this application; Figure 4 This is a schematic diagram of the structure of the condenser plate 12 provided in this application; In the diagram, 1. Base; 2. Housing; 3. Partition; 4. Integrated cooling unit; 5. Cooling plate; 6. Heat exchange unit; 7. Circulating pump; 8. Plate heat exchanger; 9. Intelligent control unit; 10. Pressure and flow sensor; 11. Venturi-diaphragm differential pressure regulating valve; 12. Condensing plate; 13. Capacitive continuous liquid level sensor; 14. Expansion tank; 15. Water replenishment device; 16. Paraffin-based phase change cold storage module; 17. Distributed fiber optic temperature sensor; 18. Variable frequency fan; 19. Spiral coil cooling coil; 20. Fiber optic grating temperature sensor array; 21. Pressure sensor. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to embodiments. It is to be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit the invention.

[0018] Example 1: In the actual operation of prefabricated box-type substations, transformer winding temperature rise exhibits significant lag and localized accumulation under conditions such as sudden load changes, harmonic impacts, or drastic fluctuations in ambient temperature. Existing air-cooled or open-loop water-cooled systems generally rely on closed-loop feedback regulation using cooling water return temperature or winding surface measuring point temperature, with response delays typically exceeding 90 seconds. When encountering a short-term impact of 120% rated load (lasting ≤3s), the temperature rise of winding hot spots is prone to exceeding the F-class insulation limit (155℃), leading to accelerated local aging and even the risk of thermal breakdown. Furthermore, traditional heat dissipation systems often have independently designed and asynchronously responding submodules (cooling, heat exchange, pumping, and control), lacking the ability to collaboratively model and proactively intervene in the multi-physics coupling link of load transients → electromagnetic losses → winding temperature rise → cooling demand → heat exchange capacity, resulting in insufficient overall system robustness and low energy efficiency.

[0019] To address the aforementioned issues, this application proposes an integrated intelligent temperature control and heat dissipation architecture centered on four-dimensional coupled prediction of electricity, heat, flow, and control. This architecture does not alter the basic structural layout of the prefabricated substation. Instead, while retaining the original base 1, enclosure 2, and partition 3 compartments (low-voltage compartment, transformer compartment, and high-voltage compartment), a highly integrated, closed-loop controllable, model-driven temperature control and heat dissipation system is constructed within the transformer compartment. Through deep collaboration between precise execution at the hardware layer and proactive decision-making at the algorithm layer, a paradigm shift from passive following to proactive adaptation is achieved.

[0020] See Figure 1 and Figure 2 This application provides an integrated intelligent temperature control and heat dissipation system for a prefabricated box-type substation, including a base 1 and a box 2 mounted on the base 1. The box 2 is divided into a low-voltage chamber, a transformer chamber, and a high-voltage chamber by a partition 3. An integrated cooling unit 4 is provided in the transformer chamber, and the integrated cooling unit 4 is connected to a closed-loop temperature control system. The system also includes an intelligent control unit 9 for coordinating the operation of each subsystem. The closed-loop temperature control system includes: a cooling plate 5, installed inside the transformer room and facing the transformer equipment, internally divided into multiple cooling chambers with air inlets between them; a heat exchange unit 6, located outside the housing 2, including a closed cooling tower or a dry heat exchanger, with a built-in variable frequency fan 18; a circulation pump 7, connected to the outlet of the cooling plate 5 and the inlet of the heat exchange unit 6, driving the cooling water circulation; an expansion tank 14 and a water replenishment device 15; and a plate heat exchanger 8, which can be optionally installed in the cooling water circuit. Among them, the intelligent control unit 9 is equipped with a load feedforward-temperature feedback dual-mode predictive controller. The controller embeds a lightweight long short-term memory neural network LSTM model. The input variables include the transformer primary current harmonic distortion rate THD, the active power change rate dP / dt, the winding temperature rise history curve measured by distributed optical fiber, and the cooling water inlet and outlet temperature difference ΔT. The output is the target flow increment ΔQ of each cooling chamber in the next 60 seconds and the pre-speed command of the variable frequency fan 18 of the heat exchange unit 6.

[0021] The integrated cooling unit 4 can refer to a functional assembly consisting of a cooling plate 5, a flow guide shroud rigidly or flexibly connected to it, and matching pipeline interfaces. Its function is to deliver the cooling medium in a directional, uniform, and controllable manner to the vicinity of the transformer's heat-generating core area, forming an attached forced convection heat exchange interface. The integrated cooling unit 4 is not mechanically fastened to the transformer body, but is positioned only by a bracket installed on the inner wall of the housing 2, allowing for ±1.5mm thermal expansion displacement. Its spatial arrangement can be adaptively adjusted according to the winding distribution characteristics of different transformer models. For example, it can be directly opposite the end of the high-voltage side winding or cover the yoke area on the iron core. This application embodiment does not make any special limitations on this.

[0022] The cooling plate 5 can be a flat metal structure, such as 6063 aluminum alloy or 304 stainless steel, integrally cast or welded. Its thickness can be set to 80mm–150mm depending on the actual heat dissipation load. Internally, it is divided into multiple cooling chambers, arranged in an array along a horizontal or vertical direction. Adjacent cooling chambers are separated by partitions 3. Air inlets are provided on the top or side walls of the partitions 3. The shape of the air inlets can be rectangular, circular, or polygonal, and their size can be set to 20mm×80mm–50mm×120mm depending on airflow requirements. Internal flow channels within the cooling chambers... It can be straight-through, serpentine, or spiral, and its cross-sectional shape can be rectangular, trapezoidal, or elliptical. The ratio of the width to the depth of the flow channel can be set to 1.2:1 to 2.5:1 according to the need to balance pressure drop and heat exchange efficiency. A distributed fiber optic temperature sensor 17 with a spatial resolution of 10 cm can be attached to the surface of the cooling plate 5 for real-time sensing of the temperature field distribution on the plate surface. The specific structural form, material selection, size parameters, and surface treatment process of the cooling plate 5 can be determined comprehensively according to the space constraints of the transformer room, the physical properties of the cooling medium, and the manufacturing cost. This application embodiment does not impose any special limitations on this.

[0023] The heat exchange unit 6 can be either a closed-loop cooling tower or a dry heat exchanger. Both are located outside the housing 2 and structurally isolated from it to avoid vibration transmission and heat radiation interference. When a closed-loop cooling tower is used, it includes a spray system, a packing layer, a water collection pan, and a built-in variable frequency fan 18. The cooling water flows in a closed loop within the coils, indirectly exchanging heat with the outside air through the pipe walls. When a dry heat exchanger is used, it is a finned tube air cooler. The cooling water flows within copper or aluminum tubes, and the external forced convection air undergoes enhanced heat exchange through the fins. The variable frequency fan 18... The type can be an EC fan or a regular AC fan, with a speed adjustment range of 0–3000 rpm. The control signal receiving end is compatible with 4–20mA analog signals or CAN bus digital commands. The installation orientation of the heat exchange unit 6 should ensure that the air inlet side avoids the exhaust vents of the housing 2 and the direct sunlight surface, and its foundation needs to be vibration-damped. The specific selection, size, heat exchange area and fan configuration of the heat exchange unit 6 can be matched to the local meteorological parameters (wet bulb temperature, wind speed), system design heat dissipation and noise limits. This application embodiment does not impose any special limitations on this.

[0024] The circulating pump 7 can be a magnetically coupled variable frequency pump or a centrifugal variable frequency pump, with a rated head of 25–45 m and a rated flow rate of 6–12 m³ / h. 3The pump body material can be cast iron, 304 stainless steel, or reinforced engineering plastic; its inlet / outlet flange nominal diameter is DN40~DN65, and the connection method conforms to ISO228-1 or GB / T7306 standards; the circulating pump 7 is connected to the outlet of the cooling plate 5 and the inlet of the heat exchange unit 6 through a stainless steel hose or seamless steel pipe. The pipeline layout should avoid air accumulation at high points and water accumulation at low points. Automatic air vent valves and drain valves are set at key nodes; the control mode of the circulating pump 7 supports constant pressure, constant flow, or PID closed-loop regulation. Its driver has an RS485 communication interface and can receive frequency commands issued by the intelligent control unit 9; the model, material, sealing form, and control protocol of the circulating pump 7 can be flexibly configured according to the system reliability requirements, media corrosivity, and field communication architecture. This application embodiment does not impose special limitations on this.

[0025] The expansion tank 14 can be a diaphragm type or a pneumatic bladder type, with a volume of 80L to 150L, ​​a pressure bearing capacity of ≥0.8MPa, and a pre-charge nitrogen pressure of 0.15 to 0.25MPa. Its connection point is located at the highest point of the pipeline on the inlet side of the circulating pump 7, used to absorb the volume expansion and contraction of the cooling water caused by temperature changes and maintain the stability of the system static pressure. The water replenishment device 15 can include a closed-loop component consisting of a float valve, an electric shut-off valve, and a liquid level switch, or it can be integrated into the body of the expansion tank 14 to realize the dual functions of automatic liquid replenishment and overpressure relief. The combination form, installation height, safety valve setting pressure, and water quality monitoring module (such as conductivity sensor) of the expansion tank 14 and the water replenishment device 15 can be designed differently according to the system scale, antifreeze requirements, and the degree of intelligent operation and maintenance. This application embodiment does not impose any special limitations on this.

[0026] The plate heat exchanger 8 is an optional component, and its installation position is located in the main water passage between the cooling plate 5 and the heat exchange unit 6, or connected in parallel to the bypass branch; its heat exchange area is 2-8m². 2 The design pressure is ≥1.0MPa, and the sealing gasket material can be EPDM or NBR, suitable for ethylene glycol aqueous solution media; the function of plate heat exchanger 8 is to realize heat exchange between transformer oil circuit and closed cooling water circuit. When the system is equipped with spiral coil cooling coil 19, the plate heat exchanger 8 becomes the core unit of oil-water heat exchange; whether it is activated, the heat exchange temperature difference set value and the secondary side medium flow direction are all dynamically decided by intelligent control unit 9 according to the operating mode; the plate material (such as AISI316 stainless steel), corrugation angle, single plate heat exchange area and pressure drop characteristics of plate heat exchanger 8 can be selected according to the heat exchange efficiency target and system resistance distribution. This application embodiment does not make special limitations on this.

[0027] The intelligent control unit 9 can be an embedded control system consisting of a PLC edge controller, a human-machine interface (HMI), a signal acquisition module, a communication gateway, and a power management module. The PLC main controller can be a Siemens S7-1200 series, Mitsubishi FX5U series, or a domestic Inovance H5U series, and should have at least 8 analog inputs (16-bit resolution), 4 analog outputs, 16 digital inputs / outputs, and dual Ethernet ports. The signal acquisition module is used to connect to distributed fiber optic temperature sensors 17, PT100 temperature sensors, electromagnetic flowmeters, pressure transmitters, and current / voltage transformers. Including Rogowski coils, etc.; the human-machine interface supports touch operation and has a three-dimensional thermal cloud map display, historical data backtracking, alarm event recording and remote maintenance interface; the intelligent control unit 9 is deployed in a dedicated control cabinet in the low-voltage room, and communicates with the transformer room via shielded twisted pair or optical fiber, and the electromagnetic interference resistance level meets the IEC61000-4-4 standard; its hardware platform selection, module expansion quantity, communication protocol (ModbusTCP, OPCUA, IEC61850) and installation location can be adapted according to the substation automation level and network security strategy, and the embodiments of this application do not impose special limitations on this.

[0028] The load-feedforward-temperature-feedback dual-mode predictive controller is the core algorithm module of the intelligent control unit 9. Its core technology lies in expanding traditional single-temperature-feedback control into a composite predictive control that integrates electrical load precursor characteristics and thermal inertia historical trajectories. The lightweight LSTM model embedded in this controller is a recurrent neural network based on time-series modeling. Its structure meets the industrial edge deployment constraints of ≤64 hidden layer nodes, <50KB weight file size, and <200ms single inference latency. The model's input variables include: transformer primary current harmonic distortion rate (THD) (reflecting additional copper losses caused by nonlinear loads), active power change rate (dP / dt) (reflecting load dynamic intensity), winding temperature rise history curves measured by distributed optical fibers (length ≥30 seconds, reflecting thermal diffusion status), and cooling water inlet / outlet temperature difference (ΔT) (reflecting current heat exchange efficiency). The model output... The output is the target flow increment ΔQ (unit: L / min) of each cooling chamber within the next 60 seconds and the pre-speed-up command (unit: rpm) of the variable frequency fan 18 of heat exchange unit 6; the training data of this LSTM model comes from the measured time series data under at least three typical operating conditions, including no-load sudden addition of 50% rated load, short-term impact overload (duration ≤3s), and day-night temperature difference cycle (-25℃→40℃). All data are processed by Z-score normalization and sliding time window segmentation; after the model is deployed, it performs one inference per second, and the output result is sent to the actuator after smoothing and filtering; the specific network structure, training hyperparameters, data annotation method and edge inference engine (such as ONNXRuntime or TensorRT) of the controller can be tailored and optimized according to the PLC computing power resources and real-time requirements. This application embodiment does not make any special limitations on this.

[0029] The core innovation of this application lies in constructing a new closed-loop temperature control paradigm that takes electrical parameter-driven thermal prediction as its logical starting point, multi-source time-series data joint modeling as its technical path, and water-air dual-channel collaborative acceleration as its execution carrier. This paradigm breaks through the response bottleneck of traditional heat dissipation systems that rely on temperature lag signals, advances control decisions to before the winding temperature rise occurs, and enables the system to proactively intervene in transient thermal disturbances.

[0030] The working process and principle of this application are as follows: When the transformer load changes, the primary side current signal is first acquired at high speed by the Rogowski coil to calculate THD and dP / dt; at the same time, the distributed optical fiber continuously monitors the evolution of the winding temperature field and extracts the historical temperature rise curve; the cooling water inlet and outlet temperature sensors synchronously output ΔT; the above four-dimensional signals are input into the LSTM model after analog-to-digital conversion; based on the learned electro-thermal coupling mapping relationship, the model predicts the incremental cooling capacity required by each cooling chamber in the next 60 seconds and outputs the corresponding ΔQ command and fan pre-speed command; the ΔQ command drives the electric three-way regulating valve to dynamically allocate the cooling water flow to different cooling chambers, realizing the precise spatial allocation of cooling resources; the pre-speed command increases the speed of the heat exchange unit 6 fan in advance, enhancing the air-side heat exchange capacity before the cooling water temperature rises significantly, compensating for the response delay caused by the thermal inertia of the water circuit; the two work together to compress the overall thermal response time and suppress the peak temperature rise of the winding.

[0031] As an optional embodiment, the specific implementation of the scheme in this application is as follows: Using a certain type ZBW-12 / 0.4–6301–ⅠA prefabricated box-type substation as the platform, the dimensions of box 2 are 2800mm × 1800mm × 2200mm, and the transformer room volume is 3.2m³. 3 It features a built-in S13-M-630 / 10 oil-immersed transformer; the cooling plate has a total size of 1600mm × 800mm × 120mm, consisting of 8 cooling chambers arranged along the height; the closed-loop circulation system has a rated flow rate of 8.5m³ / h. 3 / h, cooling water is 35% ethylene glycol aqueous solution; the intelligent control unit 9 adopts Siemens S7-1200 PLC, and expands to SM1231 module to collect 16 channels of FBG temperature, 4 channels of PT100, 2 channels of electromagnetic flowmeter and 2 channels of Rogowski coil signal; when the system detects that the primary side current dP / dt>1.2MW / s and THD>8.5%, and the slope of the winding temperature rise history curve is>0.8℃ / s for 3 consecutive seconds, the LSTM model starts the prediction process; 27 seconds The output ΔQ command (cooling chambers #3 and #4 each +1.2L / min, cooling chamber #5 +0.8L / min) and the fan pre-speed command (+420rpm) are then executed. The electric regulating valve completes the opening adjustment within 1.8 seconds, and the EC fan reaches the target speed within 3.5 seconds. Actual measurements show that under a 120% rated load impact, the highest winding temperature drops from 102.3℃ to 73.6℃, ​​the peak temperature rise decreases by 28.0%, and the thermal response time is shortened from 92 seconds to 29 seconds.

[0032] Through the above technical solution, this application achieves the following beneficial effects: By introducing the transformer primary current harmonic distortion rate (THD) and active power change rate (dP / dt) as precursor indicators of winding temperature rise, a cross-physical field mapping relationship between electrical loss and thermal response can be established, enabling the controller to issue control commands more than 27 seconds before the actual occurrence of winding temperature rise, breaking through the second-level lag bottleneck of traditional PID control; Since the LSTM model output is the target flow increment ΔQ of each cooling chamber rather than a fixed flow setpoint, the cooling water space allocation can be dynamically adjusted according to the local hot spot distribution of the winding, avoiding cold... However, resources are ineffectively circulated in the low-temperature zone, improving the heat exchange efficiency of the unit cooling medium; since the controller synchronously outputs the pre-speed command of the heat exchange unit 6 variable frequency fan 18, the air-side heat exchange capacity can be enhanced before the cooling water temperature rises, forming a dual-channel coordinated heat dissipation mechanism of water and air paths, effectively compensating for the inherent thermal inertia defects of the closed-loop system; since the LSTM model is lightweight and compressed and deployed at the edge of the PLC, the single inference delay is <200ms, thus ensuring the real-time availability of the algorithm in the industrial field, and truly transforming the feedforward predictive control from a theoretical model into an engineering-applicable technical solution.

[0033] Example 2: In one alternative implementation, such as Figure 3 As shown, this application also provides a system in which the cooling plate 5 adopts a modular and detachable structure, which is composed of multiple independent cooling units spliced ​​together in the horizontal direction; each cooling unit corresponds to a cooling chamber, and the units are connected to the water circuit through an axial floating sealing joint. The axial floating sealing joint has a built-in disc spring, which allows ±0.5mm axial displacement to compensate for thermal expansion and contraction. Each cooling unit has an integrated Venturi-diaphragm differential pressure regulating valve 11 at its water inlet. The Venturi-diaphragm differential pressure regulating valve 11 uses the kinetic energy of water to drive the deformation of the elastic diaphragm and automatically adjusts the throttling area to keep the pressure drop difference between any two adjacent cooling units constant at ≤0.8kPa. The cross-section of the internal flow channel of each cooling unit shrinks in a gradient according to the golden ratio λ=0.618.

[0034] The cooling plate 5 adopts a modular and detachable structure. This means that the cooling plate 5 is not a single cast or welded part or a single machined part, but is composed of multiple independent cooling units with the same structure and size, which are sequentially spliced ​​along the horizontal direction (i.e., parallel to the ground and perpendicular to the airflow direction). Each cooling unit can be disassembled, replaced, or maintained individually without shutting down the entire closed-loop temperature control system. This structure facilitates transportation, on-site assembly, and subsequent expansion, and also supports flexible adaptation to transformers of different power levels. The external dimensions of each cooling unit can be set according to the space and heat dissipation load requirements of the transformer room. For example, it can be 1600mm×100mm×120mm or 1400mm×120mm×110mm. This application embodiment does not make any special limitation on this. Each cooling unit corresponds to a cooling chamber, that is, the isolated cooling cavities divided within the cooling plate 5. The number of these cavities corresponds one-to-one with the number of cooling units. Each cooling chamber participates in the heat exchange process independently and there is no cross-flow between them.

[0035] The axial floating sealing joint can be a flange-type quick-connect sealing structure, comprising an outer cylinder, an inner mandrel, a disc spring assembly, and a fluororubber O-ring. The outer cylinder and the inner mandrel form an axial guide channel through a precision sliding fit. The disc spring assembly is pre-loaded and installed in a groove at the end of the outer cylinder, providing an initial preload of 120N to maintain a sealed contact without thermal deformation. When the cooling unit undergoes axial thermal expansion due to changes in ambient temperature or operating temperature rise, the disc spring is compressed, allowing the inner mandrel to freely expand and contract relative to the outer cylinder within a range of ±0.5mm, thereby releasing thermal stress and preventing micro-cracks or permanent deformation caused by hard compression of the sealing surface. The sealing performance of this joint has been verified by helium mass spectrometry leak detection, with a leakage rate ≤1×10⁻⁶. -9 Pa·m 3 Its functional positioning is as a mechanical-fluid coupling interface for water circuit connection between cooling units. While ensuring pressure resistance (rated working pressure 0.35MPa), it actively absorbs dimensional deviations caused by thermal expansion and contraction, preventing interface leakage, vibration fatigue cracking, or local distortion of the flow channel caused by rigid connection. The connection method between this connector and the cooling unit body can be ISO228-1G1. The internal thread can be locked with anaerobic adhesive, or it can be a clamp-type quick-release structure. This application does not make any special limitation on this.

[0036] The Venturi-diaphragm type differential pressure regulating valve 11 can be composed of a Venturi contraction section, an elastic diaphragm, a valve body, a return spring, and a pressure feedback chamber. The Venturi contraction section is made of 316L stainless steel, with an inlet diameter D1=42mm, a throat diameter D2=26mm, a contraction angle α=22°, a diffusion angle β=7°, and a throat length L. =18mm; the elastic diaphragm is made of fluororubber (FKM) with a thickness of t=0.8mm, an effective diameter of Φ=30mm, and a central opening of φ6mm; the valve body is made of 6061-T6 aluminum alloy by die casting, with an anodized inner surface; the pressure feedback chamber is connected to the Venturi throat through a φ1.2mm damping orifice to form a dynamic differential pressure sensing loop; the function of this valve is to utilize the kinetic energy-static pressure conversion effect generated when cooling water flows through the Venturi section, using the negative pressure at the throat as a control signal to drive the diaphragm to move, thereby adjusting the flow cross-sectional area in real time; its naming is based on the Venturi representing the differential pressure generation principle, the diaphragm representing the actuator, and the differential pressure stabilization representing the control target; The valve is integrated with the cooling unit inlet, directly embedded in the inlet flange end face of the cooling unit body, forming an inseparable functional unit. It works in conjunction with the axial floating seal joint: the joint ensures long-term reliable sealing of the interface, while the valve further ensures the consistency of hydraulic conditions across all units while maintaining a tight seal. When the flow resistance of any cooling unit increases due to manufacturing tolerances or scaling, the negative pressure at its upstream venturi throat increases, causing the diaphragm to bulge towards the throat, reducing the flow area and thus increasing the inlet pressure drop of that unit. This forces a redistribution of flow, ultimately clamping the pressure drop difference between any two adjacent cooling units within the range of 0.78–0.82 kPa. This valve can operate at flow rates from 2 to 12 m³ / h. 3 It achieves a steady flow accuracy of ±1.3% within the range of / h, and requires no external power supply, communication or control signal, and is a passive adaptive regulating device; its alternative can be a micro servo valve driven by piezoelectric ceramic, or a dual-chamber diaphragm differential pressure balancing valve, but requires additional power supply and controller, which is not specifically limited in this application.

[0037] The cross-sectional area of ​​the internal flow channels of each cooling unit decreases gradually according to the golden ratio λ=0.618. This can be interpreted as the cross-sectional area of ​​the cooling water flow path within the cooling unit decreasing geometrically along the water flow direction, with the ratio of the areas of two adjacent cross-sections always equal to 0.618. For example, if the inlet cross-sectional area is S0, then after the first contraction section, it becomes S1=S0×0.618, and after the second section, it becomes S2=S1×0.618=S0×0.618. 2 And so on; the functional positioning of this gradient contraction structure is to suppress boundary layer separation and vortex shedding phenomena in the contraction section of the flow field, making the velocity increase process smoother and the pressure recovery more sufficient; CFD simulation shows that at a Reynolds number Re=3.5×10 4Under operating conditions, compared to a straight pipe of equal diameter or a single-stage abrupt contraction structure, this golden ratio flow channel can reduce the local resistance coefficient by 12%, reduce turbulent kinetic energy dissipation by 19%, and decrease the flow field uniformity index (defined as the standard deviation of the outlet section velocity / average velocity) from 0.41 to 0.23. Its working path is as follows: the gradual contraction at each stage makes the acceleration of fluid particles continuously controllable, avoiding the adverse pressure gradient caused by the sharp bending of the streamline, thereby maintaining the attached flow state. This structure forms an upstream and downstream synergy with the Venturi-diaphragm differential pressure stabilizing valve 11: the former optimizes the internal flow state of the unit, and the latter balances the hydraulic distribution among multiple units. The cross-sectional shape of the flow channel can be rectangular, elliptical, or rounded rectangle, and its aspect ratio and radius of curvature can be set according to the shape of the cooling unit and the manufacturing process. This application embodiment does not impose any special limitations on this.

[0038] In actual operation, each independent cooling unit first completes mechanical connection and water circuit connection through axial floating sealing joints. After the system starts and water pressure is established, cooling water enters the inlet of each unit, triggering the action of the Venturi-diaphragm differential pressure regulating valve 11. If a unit has slightly high flow resistance due to slight scaling or assembly error, the negative pressure at its Venturi throat increases, the diaphragm bulges under pressure, and the throttling area automatically decreases, thereby increasing the pressure loss of the unit itself and prompting other units to divert and supplement it, ultimately achieving high synchronization of flow rates of all cooling units under the same inlet total pressure. At the same time, the cooling water flows stably along the golden ratio gradient contraction channel inside each unit, the boundary layer develops fully, and eddy current generation is suppressed, ensuring efficient convective heat transfer between the unit volume of cooling water and the transformer surface. This collaborative mechanism does not rely on external control commands and is driven entirely by the fluid's own dynamic characteristics, with a response time of milliseconds, significantly improving the system robustness and water distribution accuracy under the modular structure.

[0039] As an optional embodiment, the specific implementation of the scheme in this application is as follows: In the ZBW-12 / 0.4–6301–IA type prefabricated box-type substation, the cooling plate 5 is composed of 8 independent cooling units spliced ​​horizontally, with a single unit size of 1600mm×100mm×120mm; the units are connected by an axial floating sealing joint with a nominal diameter of DN40 and a built-in disc spring preload of 120N, which maintains the sealing integrity after 1000 thermal cycles; each cooling unit's inlet integrates a Venturi-diaphragm type differential pressure stabilizing valve 11, and bench tests show that at 2.5~11.8m 3 Within the flow rate range of / h, the pressure drop difference between any two adjacent units is stably maintained at 0.79 to 0.81 kPa; the internal flow channel of the cooling unit undergoes a five-stage gradient contraction according to λ=0.618, with an inlet cross-section of rectangular (100mm×15mm) and a final stage cross-section of 61.8mm×9.3mm; during 72 hours of load operation, the measured flow deviation of the eight cooling chambers is -3.6% to +3.7%, which meets the requirements of index b in this application.

[0040] Through the above technical solutions, this application achieves the following: Due to the use of an axial floating sealing joint, the ±0.5mm axial displacement of the cooling unit caused by thermal expansion and contraction can be absorbed, avoiding micro-leakage and sealing failure at the interface; Due to the integrated Venturi-diaphragm differential pressure regulating valve 11, the throttling area can be automatically adjusted without external energy or control signals, forcibly clamping the pressure difference between adjacent cooling units within the range of ≤0.8kPa, thereby ensuring the consistency of flow distribution in each cooling chamber under dynamic operating conditions; Due to the gradient contraction of the internal flow channels of the cooling unit according to the golden ratio λ=0.618, turbulent separation can be suppressed, the local resistance coefficient reduced, and the flow field stability and heat transfer uniformity improved; The synergistic effect of these three factors enables the modular cooling plate 5 to maintain a flow deviation within ±3.7% during long-term operation, providing a high-fidelity execution basis for the ΔQ command output by the LSTM model, fundamentally solving the problem of fluid controllability degradation caused by the structural discretization of the integrated cooling unit 4.

[0041] Example 3: In one alternative embodiment, such as Figure 4 As shown, this application also provides a multi-layer protective structure at the air inlet of the cooling plate 5, wherein the surface of the condenser plate 12 is divided into three functional areas along the airflow direction: the air inlet side 1 / 3 area is coated with a fluorosilane superhydrophobic coating with a contact angle >160° and a roll-off angle <3°; the middle 1 / 3 area is laser microtextured to form a hydrophilic area with a micro-pit array diameter of 20μm and a depth of 5μm; and the air outlet side 1 / 3 area is coated with a graphene-titanium dioxide photocatalytic hydrophobic layer. A capacitive continuous liquid level sensor 13 is connected to the bottom of the condenser plate 12. The measurement resolution is 0.1 mm. Its output signal is linked to a micro drainage pump to achieve precise drainage at the drop level.

[0042] The condenser plate 12 is a 6063 aluminum alloy flat plate with dimensions of 1200mm×600mm×3mm and a surface roughness Ra controlled at 0.8±0.1μm. As the core carrier of the multi-layer protective structure, the condenser plate 12 is positioned to coordinate and regulate the entire process of water vapor condensation, spreading, guiding and discharging in the airflow, so as to avoid condensate water from stagnating, bridging or accumulating in the air inlet channel, thereby ensuring continuous and stable air intake efficiency and air cleanliness. It is fixed to the cooling plate 5 body by stainless steel quick-release buckles, can be disassembled and replaced, and has an installation tilt angle of 15° to adapt to the airflow direction and enhance the gravity drainage path.

[0043] The fluorosilane superhydrophobic coating on the air inlet side 1 / 3 area uses Dow Corning SYLGARD technology. A hydrophobic coating formulated with 1.5 wt% nano-SiO2 was sprayed and cured at 120°C for 30 min, resulting in a film thickness of 12 ± 2 μm. This coating is a stable superhydrophobic surface with a contact angle of 162° ± 2° and a roll-off angle of 2.1° ± 0.3°. Its name is based on the Cassie-Baxter state constructed through the synergistic effect of a low surface energy fluorosilane framework and a nano-rough structure, which allows condensed water droplets to bounce off in a spherical shape, thus efficiently intercepting and instantly rolling off condensed water in the initial airflow section, preventing droplets from being carried into subsequent flow channels by the airflow. The coating and the central hydrophilic zone are separated by a high-precision mask, with an overlap error ≤ 0.1 mm. Its function is to provide rapid flow guidance, forming a directional water flow guidance link driven by a pressure gradient with the central hydrophilic zone.

[0044] The central third region's laser-textured hydrophilic area is an array of micro-pits formed on an aluminum alloy substrate by a picosecond laser (wavelength 1064nm, pulse width 12ps, frequency 200kHz). The micro-pits have a diameter of 20±0.5μm, a depth of 5±0.3μm, and a spacing of 25μm, arranged in an orthogonal periodic pattern. This structure actively adsorbs and spreads a condensate film in the opposite direction of the airflow through capillary action, inhibiting its lateral expansion to form a water film bridge, thereby maintaining the effective flow cross-section of the air inlet channel. It forms a wettability step difference with the superhydrophobic region on the air inlet side, constituting a wetting dynamic chain from repulsion to capture to directional guidance. The functional positioning of this region is to achieve controllable spreading and axial flow of condensate, forming a front-end water film blocking mechanism in conjunction with the air inlet side coating, and forming a closed-loop moisture management path together with the photocatalytic hydrophobic layer on the air outlet side.

[0045] The graphene-titanium dioxide photocatalytic hydrophobic layer in the 1 / 3 area of ​​the air outlet side is a composite slurry coating prepared by spin coating, wherein the graphene content is 0.8wt%, the P25TiO2 particle size is 25nm, the film thickness is 8±1μm, and it is subjected to 365nm ultraviolet light (50mW / cm²). 2 The coating was obtained by curing for 2 hours under UV irradiation. It is a hydrophobic surface with a contact angle of 156°±3° and a degradation rate of methyl orange solution of >92% / h under UV irradiation. The name is based on the fact that graphene improves the carrier separation efficiency and TiO2 provides photocatalytic activity, which synergistically achieves in-situ decomposition of organic pollutants and long-term maintenance of surface hydrophobicity. The coating plays a role in self-cleaning and maintaining the hydrophobic performance, preventing the deposition of microorganisms and dust, and ensuring the long-term synergistic effectiveness of the three functional regions. It is also isolated from the central hydrophilic region by a mask, and the interface transition is smooth without macroscopic steps.

[0046] A capacitive continuous liquid level sensor 13 is embedded in an interdigitated electrode structure on a flexible PCB substrate at the bottom of a condenser plate 12. The substrate is made of polyimide (0.15 mm thick). The electrode width is 0.3 mm, the gap is 0.2 mm, and the length is 50 mm. The surface is covered with a 15 μm thick hydrophobic PDMS encapsulation layer, exposing only the top sensing surface. This sensor has a measurement range of 0–20 mm, a capacitance response range of 12.5–48.7 pF, and a linearity R0. 2 The high-precision continuous liquid level sensing unit has a capacitance of 0.9992 and a resolution of 0.08mm. Its function is to replace the traditional mechanical float / electrode type liquid level switch, eliminate the threshold dead zone and response hysteresis, and realize full-range, real-time, and linear monitoring of the condensate accumulation process. The sensor and the miniature drain pump form a closed-loop control circuit. When the capacitance value changes and the corresponding liquid level rises to 0.3mm, a drain command is triggered to achieve precise drainage at the drop level. The system linkage is as follows: after the sensor output signal is collected by the PLC analog input module (SM1231, 16-bit resolution), it drives the miniature drain pump (rated voltage 24VDC, flow rate 0.8L / min, head 1.2m) to start and stop in real time, so that the condensate residence time is controlled within ≤15s, completely avoiding the risk of microbial growth and filter clogging caused by liquid accumulation.

[0047] The miniature drainage pump is a brushless DC diaphragm pump. Its inlet and outlet are connected to the water collection tank at the bottom of the condenser plate 12 via silicone hoses. The inner diameter of the hose is φ4mm, and it is located 5mm downstream of the lowest point of the condenser plate 12. The pump shares the same power supply and control bus with the capacitive sensor. The flow rate is finely adjusted by the PWM signal output by the PLC (duty cycle adjustable, 0~100%). Its function is to perform the drainage action as the terminal actuator, forming a minimum closed loop of perception-decision-execution with the sensor, supporting the physical realization of the drip-level control target.

[0048] The three functional zones work in tandem along the airflow direction: When high-temperature and high-humidity air flows through the superhydrophobic zone on the air inlet side, water vapor preferentially condenses into discrete water droplets on the low-temperature surface and quickly rolls off under the shearing force of the airflow and gravity; the tiny water droplets that have not completely detached migrate with the airflow to the hydrophilic zone in the middle, and are actively spread into a thin water film under the capillary force of the micro-pits, and are guided to the air outlet side along a preset angle; after reaching the air outlet side, the water film is continuously purified by UV light on the graphene-TiO2 coating surface, while maintaining a hydrophobic state to facilitate final detachment; throughout the process, the capacitive sensor continuously monitors the liquid level in the bottom water collection tank, and once a liquid level increase of ≥0.3mm is detected, the micro-drain pump is immediately started to discharge the accumulated condensate, with the single drainage volume controlled within the range of 0.5 to 2.0mL, achieving true water discharge upon sight and control upon dripping.

[0049] As an optional embodiment, the solution of this application is specifically implemented as follows: the system runs continuously under high humidity conditions of 25°C and 92% relative humidity. When the incoming air passes through the condenser plate 12, the superhydrophobic zone on the inlet side causes more than 92% of the condensate droplets to roll off within 0.8s. After the residual water vapor migrates to the hydrophilic zone in the middle, it forms a directional water film with a thickness of <80μm under the guidance of the micro-pit array, and flows smoothly to the outlet side at a 15° angle without any lateral bridging. The photocatalytic hydrophobic layer on the outlet side continuously decomposes the attached organic matter under natural light (illuminance ≥500lux), with a surface contact angle attenuation rate of <0.5° / 100h. The capacitive sensor outputs the liquid level signal in real time, and the PLC samples once every 200ms. When the liquid level reaches 0.32mm, the drain pump is triggered, draining 1.3mL at a time. After draining, the liquid level drops back to 0.05mm, and the entire draining cycle is 12.4s. Continuous monitoring for 72h shows that the air intake efficiency is maintained at 94.1%, the wind resistance increase is 1.3%, and the filter pressure difference growth rate is reduced by 89%, verifying the effectiveness of this multi-layer protective structure in the coordinated control of the entire condensate flow process.

[0050] Through the above technical solutions, this application achieves the following: Since the surface of the condenser plate 12 is divided into an inlet superhydrophobic region, a central hydrophilic region, and an outlet photocatalytic hydrophobic region along the airflow direction, a wetting gradient chain of repulsion-capture-purification is formed, solving the inherent contradiction that a single wetting characteristic cannot simultaneously address flow diversion, spreading, and self-cleaning; Since the contact angle of the superhydrophobic region is >160° and the roll-off angle is <3°, the central micro-pit has a diameter of 20μm and a depth of 5μm, and the coating on the outlet side has photocatalytic activity, each functional area maintains structural and chemical stability under high humidity conditions, ensuring long-term synergistic effectiveness; Since the integrated capacitive continuous liquid level sensor 13 (resolution 0.1mm) is linked to a micro-drainage pump, the threshold dead zone and response delay of the mechanical liquid level switch are eliminated, achieving precise drainage at the droplet level, reducing the condensate residence time by 92%, and decreasing the amount of microbial attachment by 93.6%, thereby significantly improving the system's adaptability and long-term operational reliability under harsh weather conditions.

[0051] Example 4: In one possible implementation, this application also provides a system in which the cooling medium used in the closed-loop temperature control system is an aqueous solution of ethylene glycol; the expansion tank 14 also integrates a paraffin-based phase change cold storage module 16 with a phase change temperature of 5°C; The circulating pump 7 is a magnetically coupled variable frequency pump with a carbon fiber reinforced PEEK impeller. The impeller flow channel incorporates a micro-heating wire mesh, and its power density is 0.8 W / cm³. 2 ; The intelligent control unit 9 is also equipped with a viscosity compensation mode: when the cooling water temperature is detected to be <-10℃ and the real-time flow rate is reduced by >15% compared with the rated value, the micro heating wire is activated to locally heat the 0.5mm liquid layer around the impeller, so that the dynamic viscosity of this area is reduced by ≥40%.

[0052] The cooling medium used in the closed-loop temperature control system is a 30%–50% volumetric ethylene glycol aqueous solution, such as 35%, 40%, or 45% ethylene glycol aqueous solution, with a freezing point as low as -30°C and a dynamic viscosity of 128 mPa·s at -30°C. This cooling medium is used to maintain the system's fluidity and heat transfer stability over a wide temperature range, and its specific concentration is determined based on the local extreme minimum ambient temperature and the system's pressure resistance. This cooling medium is chemically compatible with all contact components in the system—including 304 stainless steel piping, 8 copper brazed flow channels in the plate heat exchanger, carbon fiber reinforced PEEK impeller, and fluororubber seals—and does not cause stress corrosion or swelling failure.

[0053] The expansion tank 14 is a vertical air-bladder type expansion tank with a volume of 100L to 150L, ​​for example, 120L. Its inner liner is made of 316L stainless steel, and the inner wall is polished to Ra≤0.4μm to inhibit microbial adhesion. The paraffin-based phase change cold storage module 16 is a cylindrical package embedded in the liquid-containing area below the upper air-bladder cavity of the expansion tank 14. Its phase change material composition is n-octadecane / expanded graphite composite material (mass ratio 92:8), with a measured phase change temperature of 5.1℃±0.2℃, a phase change latent heat of 185kJ / kg, and a thermal conductivity of 1.2W / (m·K). This module absorbs and stores the excess cold energy of the cooling return water during the high-temperature period in summer and releases the cold energy before the low-temperature start-up in winter, so that the pipeline system is pre-cooled to the range of 8℃ to 12℃ before power-on, thereby shortening the time required for the system to establish stable pressure and flow. The geometric dimensions, installation position, and packaging form of this module are adaptively adjusted according to the internal space layout of the expansion tank 14.

[0054] Circulating pump 7 is a magnetically coupled variable frequency pump, such as Grundfos MAGNA340-130F or Wilo TOP-S50 / 250, with a rated head of 45m and a rated flow rate of 8.5m³ / h. 3 / h, working pressure rating PN16; its impeller material is carbon fiber reinforced polyetheretherketone (CF / PEEK), in which the mass fraction of carbon fiber is 25% to 35%, for example, 30wt%. This material has both high specific strength (≥280MPa) and low coefficient of thermal expansion (2.2×10). -6(K) and excellent chemical resistance; the impeller flow channel surface is laser-processed to form an embedded groove network, with a groove width of 0.12mm~0.18mm and a depth of 0.06mm~0.10mm, used to hold the nickel-chromium alloy micro-heating wire; the micro-heating wire mesh has a planar orthogonal arrangement structure, with a wire diameter of 0.10mm±0.01mm and a sheet resistance of 2.3Ω / cm. 2 ~2.5Ω / cm 2 The overall power density is 0.8 W / cm³. 2 ±0.05W / cm 2 This design ensures that heat is concentrated only on the 0.5mm thick boundary layer liquid film adjacent to the impeller surface, avoiding a global temperature rise in the mainstream coolant, thus achieving targeted viscosity control with minimal energy consumption. The layout density, power supply method, and insulation encapsulation process of the micro heating wires are adapted and optimized according to different impeller curvatures and flow channel hydraulic characteristics.

[0055] The viscosity compensation mode configured in the intelligent control unit 9 is a closed-loop triggering logic based on multi-parameter collaborative criteria: its input signals include the cooling water return temperature T collected by the PT100 temperature sensor (Class A accuracy, ±0.15℃) and the real-time volumetric flow rate Q output by the electromagnetic flow meter (accuracy ±0.5%FS); when the PLC edge controller detects for 2 seconds that T < -10℃ and Q < 0.85 × Q (Q) When the system's rated flow rate is reached, the viscosity compensation subroutine is automatically activated. This subroutine outputs a PWM control signal (duty cycle 30%~40%, frequency 10Hz) to drive the micro-heating wire, maintaining the temperature of the 0.5mm liquid layer around the impeller at -5℃±1℃. The measured dynamic viscosity of the ethylene glycol aqueous solution in this region decreased from 128mPa·s to ≤76mPa·s, a reduction of ≥40.6%. This mode is only enabled under low-temperature, high-constraint conditions and is completely decoupled from the system's conventional temperature control logic, without affecting the independence and stability of other operating modes. Its trigger threshold, response delay, heating duration, and exit conditions are configured online through the HMI interface.

[0056] The working process of viscosity compensation mode is as follows: In the initial stage of cold start-up in an environment of -30℃, the high viscosity of the cooling water causes the pump outlet pressure to rise and the actual flow rate to remain lower than the set value. After the intelligent control unit 9 identifies this abnormal state, it does not take traditional power-enhancing measures such as increasing the pump speed or increasing the valve opening, but instead precisely activates the micro heating wire to apply local thermal excitation only to the boundary layer region at the impeller flow channel outlet where flow separation is most likely to occur. Since this thin layer of liquid is directly subjected to the shearing action of the impeller, its viscosity decreases, which directly improves fluid adhesion and energy transfer efficiency, restoring the pump efficiency to 92.4% and the flow deviation to within ±3.7%. At the same time, the PCM module releases the previously stored cold energy, reducing the overall temperature rise rate of the pipeline and avoiding local overheating that could lead to ethylene glycol degradation. The synergistic effect of the two allows the system to complete the entire process of rapid start-up and shutdown from static evacuation to full-load operation without the need for external heat source intervention.

[0057] As an optional embodiment, the specific implementation of the scheme in this application is as follows: In a certain type ZBW-12 / 0.4–6301–ⅠA prefabricated box-type substation, the closed-loop circulating temperature control system uses a 45% volume concentration ethylene glycol aqueous solution as the cooling medium; the expansion tank 14 is a 120L vertical airbag tank, with an internal n-octadecane / expanded graphite PCM module with a diameter of 180mm and a height of 300mm; the circulating pump 7 is a Grundfos MAGNA340-130F magnetic coupling pump, the impeller is made of 30wt% carbon fiber reinforced PEEK material, and the flow channel is embedded with a nickel-chromium micro heating wire mesh (power density 0.8W / cm²). 2 When the ambient temperature suddenly drops to -30℃, the system enters a power-off shutdown state. The PLC detects that the return water temperature sensor reading is -10.3℃ and the flow meter feedback value is 7.12m³. 3 / h (lower than the rated value of 8.5m) 3 When the viscosity reaches 16.2% ( / h), the viscosity compensation mode is immediately activated; the micro-heating wire is energized at a 35% duty cycle, and after 2.3 seconds, the impeller outlet boundary layer temperature rises to -4.8℃, and the viscosity drops to 75.6 mPa·s; combined with the pre-stored cold energy released by the PCM, the system completes the entire process of liquid injection, venting, pressure building, and flow stabilization within 4 minutes and 18 seconds, with the cooling water pressure reaching 0.35 MPa and the flow rate stabilizing at 8.47 m³ / h. 3 / h, meeting the performance target of ≤5min startup time in winter mode.

[0058] Through the above technical solutions, this application achieves the following: A paraffin-based phase change cold storage module 16 is integrated into the expansion tank 14, enabling the spatial and temporal transfer of cold energy during periods of surplus cold energy in summer, providing a pre-temperature control foundation for low-temperature startup in winter; the impeller of the circulating pump 7 is made of carbon fiber reinforced PEEK material with an embedded micro-heating wire mesh, allowing for targeted control of the impeller boundary layer liquid film viscosity without altering the state of the mainstream cooling medium; the intelligent control unit 9 is equipped with a viscosity compensation mode, triggering local heating based on a joint criterion of temperature and flow rate, avoiding the high energy consumption and thermal stress risks associated with traditional global heating; the synergistic effect of these three components resolves the fundamental contradiction between antifreeze reliability and system energy efficiency in low-temperature environments, enabling the integrated intelligent temperature control and heat dissipation system to operate continuously in all climates.

[0059] Example 5: In yet another alternative embodiment, such as Figure 1 , Figure 2 As shown, this application also provides that the structural parameters of the LSTM model meet the following requirements: the number of hidden layer nodes ≤ 64, the training dataset is derived from measured time series data under at least three typical working conditions, including: sudden addition of 50% rated load under no-load conditions, short-term impact overload duration ≤ 3s, and day-night temperature difference cycle -25℃→40℃; the model weight file size is < 50KB, it is deployed in a PLC edge controller, and the single inference latency is < 200ms.

[0060] The number of hidden layer nodes in the LSTM model can be 48, 56, or 64, with the specific value configured based on the available memory capacity and real-time margin of the PLC controller. When using a Siemens S7-1200 series PLC (with built-in RAM ≥ 1MB and a main frequency ≥ 1GHz), 48 nodes are an option, ensuring model fitting capability while reserving sufficient resources for other control tasks. The upper limit of 64 nodes is based on the identification results of the order of the nonlinear dynamic system of winding temperature rise—via MATLAB System. Modeling the measured temperature rise response of 12 types of transformers using mIdentificationToolbox shows that the equivalent state space dimension does not exceed 6, corresponding to a theoretical upper limit of 6×8=48 for the number of hidden layer nodes in the LSTM. After leaving 25% redundancy, the number is determined to be 64. Although the number of nodes is less than 48, the volume can be further compressed, but it will lead to an increase in prediction deviation under short-term impact overload conditions (the measured MAE increases from 0.17℃ to 0.29℃). Therefore, this application does not impose a lower limit on the number of hidden layer nodes, but only uses ≤64 as an upper limit constraint.

[0061] The training dataset is derived from measured time-series data under at least three typical operating conditions. These can be any combination of the following: sudden application of 50% rated load under no-load conditions, short-term impact overload duration ≤3s, and diurnal temperature cycle from -25℃ to 40℃. Alternatively, it can be a mixed sequence containing all three conditions. The sudden application of 50% rated load under no-load conditions is used to characterize the thermal inertial response process induced by steady-state load transitions, with a sampling duration of at least 120 minutes and a sampling frequency of at least 1Hz. The short-term impact overload duration ≤3s is used to characterize transient thermal stress impacts, requiring the pulse width to be strictly controlled within the range of 2.0–3.0s and a repetition interval ≥60. The cumulative number of valid impact samples collected shall not be less than 200; the diurnal temperature cycle of -25℃ to 40℃ is used to verify the robustness of the model under drastic changes in environmental boundary conditions. The temperature rise and fall rate shall be controlled at ±5℃ every 2 hours in accordance with GB / T2423.1–2008, and the entire process shall last for 48 hours; these three types of operating conditions cover the most representative electro-thermal coupling dynamic scenarios in substation operation. The joint training of their data can reduce the extrapolation error of the model under unknown operating conditions by 42% (compared to training under a single operating condition); the embodiments of this application do not impose additional limitations on the number of operating conditions and the specific parameter range, but only require the listed three types as the minimum coverage.

[0062] The model weight file size is <50KB, and can be 47.3KB, 42.6KB, or 49.8KB. This is achieved through a three-layer collaborative compression: the first layer is structural pruning, which removes connection weights in the LSTM gate unit whose contribution is below a threshold (0.005); the second layer is INT8 quantization, which maps the FP32 weights to an 8-bit integer field and uses a channel-wise asymmetric quantization strategy to preserve key feature scales; the third layer is layer fusion and operator optimization, which merges the BatchNorm and Linear layers in the ONNX format to eliminate redundant calculations. After compilation with TensorRT8.6, the weight file size stably converges to 47.3±1.2KB. This size constraint ensures that the model can be completely stored in the FAT32 partition of a standard PLC SD card (≥1GB), and the loading time is <150ms, without affecting the cold start process. This application does not limit the compression method, as long as the final file size meets the requirement of <50KB.

[0063] The model is deployed in a PLC edge controller, which can be a Siemens S7-1200, Schneider M262, or Huichuan H5U series PLC. Its hardware platform must have: dual-core ARM Cortex-A7 or higher architecture, ≥512MB DDR3 RAM, and support for the CODESYS V3.5 runtime environment. The PLC synchronously acquires four types of input signals at a frequency of 10Hz: THD, dP / dt, ΔT, and winding temperature rise history curve through the SM1231 analog input module, and sends them to the model after Z-score standardization. The model inference results (8-dimensional ΔQ vector and 1-dimensional fan pre-speed command) drive the electric regulating valve and EC fan driver through the SM1223 digital output module. This application embodiment does not limit the PLC brand and model, only requiring that it meets the basic specifications of IEC61131-3 standard regarding real-time task cycle (≤200ms) and memory management.

[0064] The single inference latency is <200ms, which can be 186ms, 192ms, or 199ms. The measured value depends on the PLC main frequency, the model input window length, and the degree of hardware acceleration support. In this embodiment, the input window is fixed as 30 seconds of historical data (i.e., 30 time steps), and each time step contains 6-dimensional features, forming a total of 30×6=180-dimensional input tensor. On the Intel Core i5-1135G7@2.4GHz platform (simulating PLC edge computing power), the measured average inference latency is 186ms, with a standard deviation of ±3.2ms, executed by calling the TensorRT engine via ONNXRuntime. This latency meets the hard requirement of closed-loop control for the timeliness of feedforward instructions—after a 120% impact overload occurs, the model can issue the first flow increase instruction within 27s, which is 65s earlier than traditional PID feedback control. This embodiment does not limit the latency measurement method, but only uses the time interval from the completion of input sampling to the issuance of the output instruction as the standard.

[0065] Specifically, in actual operation, the LSTM model first triggers a data acquisition cycle once per second by a high-speed timer of the PLC, synchronously acquiring the current harmonic distortion rate (THD), the rate of change of active power (dP / dt), the temperature difference between the inlet and outlet of the cooling water (ΔT), the ambient temperature, the real-time load rate, and the historical temperature rise curve of the winding fitted by the distributed optical fiber. All raw signals are truncated through a sliding window to extract the most recent 30-second sequence and concatenated into a 6-dimensional × 30-step input tensor. The tensor is normalized by preset standardization parameters and then fed into the loaded lightweight LSTM model. After the model completes a single forward inference... It outputs 8 channels of target flow increment ΔQ (unit: L / min) for the cooling chamber and 1 channel of EC fan pre-speed command (unit: rpm); each ΔQ command is converted into a duty cycle signal by the SMCITV2050 electric three-way valve driver, which controls the opening of the water inlet branch of the corresponding cooling unit; the fan command is sent to the EC fan built-in controller through the ModbusTCP protocol to realize the speed advance adjustment; the whole process is completed within a single PLC scan cycle (200ms), without occupying additional interrupt resources, ensuring that it runs in parallel with other protection logic without conflict.

[0066] As an optional embodiment, the specific implementation of the scheme in this application is as follows: In a certain type ZBW-12 / 0.4–6301–ⅠA prefabricated box-type substation, the intelligent control unit 9 adopts a Siemens S7-1215CDC / DC / DC PLC, and expands the SM1231 module to collect 16 channels of FBG temperature, 4 channels of PT100, and 2 channels of Rogowski coil current signals; the LSTM model is fixed in ONNX format on an SD card and loaded by calling the ONNXRuntimeAPI via CODESYSV3.5; when the system detects a sudden increase in primary side current and THD > 8.5%, When dP / dt > 12kW / s, it is determined to be a 50% load increase under no-load conditions; the model immediately starts inference, and after 30 seconds of historical data input, it outputs 8 ΔQ commands (range +0.8 to +2.4L / min) and a fan pre-speed command (+180rpm) within 186ms; the electric valve responds within 120ms, the fan speed increases to the target value within 210ms, and the cooling water flow rate increases to a new steady state within 2.3s, effectively suppressing the winding temperature rise rate; the entire process requires no manual intervention and does not rely on SCADA host computer participation, and is completed entirely by the PLC local closed loop.

[0067] Through the above technical solutions, this application achieves the following: The structural constraint of ≤64 hidden layer nodes reduces model complexity and computational load, making LSTM feasible for native deployment on resource-constrained PLCs; the training data covers three typical operating conditions—no-load surge, short-term impact, and day-night temperature difference—improving the model's generalization adaptability to real-world power grid operating scenarios; the model weight file size is <50KB, adapting to industrial PLC embedded storage media and avoiding reliability and cost risks associated with external AI acceleration cards; and the single inference latency is <200ms, meeting the IEC61131-3 requirements for the cycle of real-time control tasks, ensuring the actual effectiveness of the load feedforward function. The synergistic effect of these four aspects solves the engineering bottleneck of existing technologies where intelligent algorithms are theoretical and difficult to implement, transforming LSTM predictive control from a laboratory model into a mass-producible, verifiable, and maintainable industrial-grade functional module.

[0068] Example 6: In an optional embodiment, this application also provides a distributed optical fiber temperature sensor 17 attached to the surface of the cooling plate 5, with a spatial resolution of 10 cm and a temperature measurement accuracy of ±0.5℃; the data from the optical fiber sensor, after being fused by wavelet denoising and Kalman filtering, is used as the input source for the winding temperature rise history curve in the LSTM model.

[0069] The distributed fiber optic temperature sensor 17 is a Raman scattering (DTS) distributed fiber optic temperature measurement system. Its sensing cable uses multimode silica fiber (core diameter 62.5μm, cladding diameter 125μm), covered with a high-temperature resistant fluorinated ethylene propylene (FEP) sheath (long-term temperature resistance -40℃ to +200℃). The fiber is laid along the surface of the cooling plate 5 in a serpentine path, with the spacing between adjacent measuring points strictly controlled at 10cm. A 20cm redundancy length is reserved at both ends for splicing and stress release. The spatial resolution of 10cm for this fiber optic sensor refers to the length of the smallest independent temperature measuring unit that can be distinguished in continuous temperature measurement mode, corresponding to a temperature sampling point density of 10 per meter. The cooling plate 5 (1600mm×800mm) has 128 effective temperature measurement nodes. The temperature measurement accuracy of ±0.5℃ is obtained by calibration according to the IEC61757-1:2018 standard under the conditions of ambient temperature of 25℃, fiber optic bending radius ≥30mm and no strong electromagnetic interference. In actual operation, this accuracy can be compensated for online drift by periodically referring to the PT100 reference point. The sensor does not directly contact the transformer winding, but indirectly reflects the surface temperature distribution of the winding through the heat conduction of the metal substrate of the cooling plate 5. Its response time τ≤3.2s (90% step response) meets the requirements of the LSTM model for capturing the dynamic process of temperature rise.

[0070] The temperature-sensing optical cable of the distributed optical fiber temperature sensor 17 is evenly attached to the inner side of the back plate of the cooling plate 5 using thermally conductive silicone (thermal conductivity ≥1.8W / (m·K)). The attachment area covers the section directly opposite the high-voltage and low-voltage windings of the transformer. The cable laying path avoids structural abrupt changes such as bolt mounting holes, sealing joints, and flow channel bosses to avoid micro-bending losses caused by local stress concentration. Flexible stainless steel cable ties (6mm wide and 0.5mm thick) are set at the four corners and the midpoint of the long side of the cooling plate 5 for mechanical restraint to prevent the optical cable from slipping due to thermal expansion and contraction. All splices are encapsulated in an IP67-rated optical fiber splice box and have built-in miniature temperature and humidity sensors for status monitoring.

[0071] The wavelet denoising process uses the Daubechies4 (db4) wavelet basis with 5 decomposition layers. The threshold selection strategy is the SureShrink adaptive thresholding method, which performs soft thresholding on high-frequency detail coefficients and retains low-frequency approximation coefficients as the backbone of the denoised signal. This processing can effectively suppress broadband noise (frequency range 1kHz~5MHz) introduced by the vibration of the housing 2, electromagnetic pulses from switching operations, and power supply ripple, improving the signal-to-noise ratio of the original signal from a typical 18dB to ≥32dB.

[0072] The Kalman filter is a first-order linear discrete Kalman filter, whose system state equation is modeled as the winding thermal inertia transfer function:

[0073] in, Estimate the temperature at time k. The temperature value at the k-th sampling point after wavelet denoising. The thermal time constant decay factor, The sampling period is The measured thermal time constant of the winding-cooling plate 5 coupling (calibrated using an infrared thermal imager); observed noise covariance. Set as Process noise covariance Set as Initial estimation error covariance The filter suppresses low-frequency drift (such as day-night temperature drift) while maintaining the ability to track temperature rise jumps. The measured steady-state temperature jitter is < ±0.12℃ and the dynamic step response overshoot is <2.3%.

[0074] Wavelet denoising and Kalman filtering are deployed in a cascaded manner in the edge computing module of the intelligent control unit 9: the original fiber optic data is first denoised in real time by the wavelet transform module accelerated by FPGA hardware (single channel delay < 8ms), and the output sequence is sent to the Kalman filtering algorithm running on the PLC's built-in ARM Cortex-M7 core (single point operation time < 1.2ms). Finally, the fused output temperature sequence is updated at a frequency of 1Hz and cached as a sliding window of 180 points (corresponding to 180 seconds of history), which serves as the data source for the winding temperature rise history curve dimension in the input feature vector of the LSTM model. This fusion process does not change the original temperature dimensions and physical meaning, but only improves the temporal continuity, spatial consistency and anti-disturbance robustness of the data.

[0075] The process of the distributed fiber optic temperature sensor 17 working in conjunction with the wavelet-Kalman fusion filter is as follows: When a sudden change in transformer load causes a local temperature rise in the winding, heat is conducted to the fiber optic laying surface through the tank wall and the substrate of the cooling plate 5. The 10cm spatial resolution ensures that the hot spot interval of 8–12cm between adjacent winding discs can be distinguished. Wavelet denoising quickly eliminates transient electromagnetic interference pulses (typical amplitude ±3.5℃, duration <50ms) generated by the opening and closing of the circuit breaker, avoiding false triggering of LSTM abnormal prediction. Kalman filtering performs state smoothing and trend prediction on the denoised sequence based on the thermal inertia model, identifying the temperature rise acceleration change before the cooling water flow responds, providing LSTM with an earlier, more stable, and more physically consistent temperature rise evolution trajectory.

[0076] As an optional embodiment, the specific implementation of the scheme in this application is as follows: In the ZBW-12 / 0.4–6301–ⅠA box-type substation of embodiment 1, DTS optical fiber (model SENSiSDTS-128) is laid along the back plate of cooling plate 5, covering an area of ​​1600mm×600mm (corresponding to the high-voltage winding projection area), with a total of 128 measuring points; the optical fiber signal is connected to a dedicated DTS interface card (model S7-DTS-16) extended by the Siemens SM1231-8RTD module, and transmitted via the built-in F The PGA completes the db4 wavelet 5-level decomposition and SureShrink thresholding. The filtered temperature sequence is then processed by the PLCCPU1215C using the embedded KalmanFilter_V1.2 function library to perform Kalman recursion, outputting a 180-point sliding temperature array. This array is refreshed every second and serves as the data source for the winding temperature rise history curve field in the LSTM model input X(t). In the 120% short-time impact load test, this fusion processing enables the LSTM to accurately predict the peak winding temperature rise ΔT at t+60s. 60 The prediction absolute error (MAE) decreased from 0.61℃ to 0.19℃, significantly improving the accuracy of feedforward commands.

[0077] Through the above technical solution, this application achieves the following: a distributed optical fiber temperature sensor 17 is attached to the surface of the cooling plate 5 with a spatial resolution of 10 cm to construct a fine temperature field that reflects the true heat distribution on the winding surface, replacing the temperature rise characterization distortion caused by traditional single-point temperature measurement; wavelet denoising and Kalman filtering are applied sequentially to the original optical fiber data to simultaneously suppress high-frequency electromagnetic noise and low-frequency thermal drift, so that the winding temperature rise history curve input to the LSTM model has high temporal fidelity and high spatial representativeness; the LSTM model accurately predicts the temperature rise trend in the next 60 seconds based on high-quality spatiotemporal sequences, supporting the load feedforward-temperature feedback dual-mode predictive controller to achieve precise and timely cooling power regulation.

[0078] Example 7: In an optional embodiment, this application also provides that the variable frequency fan 18 of the heat exchange unit 6 is an EC fan, and its speed control strategy includes: basic mode: closed-loop adjustment based on the cooling water return temperature; feedforward enhancement mode: receiving the pre-speed-up command output by the LSTM model, and increasing the fan speed before the return water temperature rises; natural cooling intervention mode: when the ambient wet-bulb temperature is lower than the cooling water set temperature and ΔT>3K, the circulation pump 7 is turned off, and only the fan is used for air-water natural convection heat exchange.

[0079] Among them, the EC fan is a brushless DC external rotor fan with a rated air volume of 12000 m³ / h. 3 The fan operates at a speed of 1000 rpm, with a total pressure of 450 Pa, a motor efficiency of ≥85%, and a speed regulation resolution of 1 rpm. It supports 0-10V analog input and Modbus RTU dual-protocol control. The impeller is made of aerospace-grade polycarbonate through a single injection molding process, with 9 blades and a 32° back tilt angle. The hub and motor shaft are interference-fitted and secured with anaerobic adhesive to ensure dynamic balance stability under long-term vibration conditions. The EC fan is mounted via a flange at the top of the heat exchange unit 6 housing with a pre-reserved interface. The interface seal uses a fluororubber O-ring (Shore hardness 70A), and the compression ratio is controlled at 25% ± 3%, meeting the IP55 protection level requirements; in this application system, the EC fan undertakes the task of forced convection heat exchange between the cooling water circuit and the external environment. Its air volume output directly determines the upper limit of the heat dissipation capacity of the heat exchange unit 6, and forms a coordinated control relationship with the flow rate of each cooling chamber of the cooling plate 5. When the LSTM model outputs the ΔQ command to increase the flow rate of a certain cooling chamber, the feedforward enhancement mode synchronously increases the speed of the EC fan, so that the newly added heat can be carried out in time, avoiding the continuous accumulation of cooling water temperature; conversely, when running under low load steady state, the basic mode maintains a lower speed to reduce noise and energy consumption.

[0080] The basic mode can refer to: the PLC acquires the signal from the cooling water return temperature sensor (PT100, Class A, measurement range 0~80℃, accuracy ±0.15℃) in real time, and outputs a 0~10V voltage signal to the EC fan driver after PID algorithm calculation. The PID parameters are tuned as follows: proportional band P=8℃, integral time Ti=120s, derivative time Td=8s. This mode forms the lowest-level temperature negative feedback closed loop, ensuring that the system still has basic temperature control capability when there is no feedforward instruction or model failure. Its action path is: return water temperature rises → PID output voltage rises → EC fan speed increases → air volume increases → heat exchange intensity increases → return water temperature falls back. The actual measured response time of the entire closed loop is 42s (from temperature exceeding the limit to air volume reaching 90% of the target value). This mode does not rely on the LSTM model or external communication link when running independently, and has fail-safe characteristics.

[0081] The feedforward enhancement mode works as follows: The EC fan receives a pre-speed-up command signal from the intelligent control unit 9. This command is output by the lightweight LSTM model defined in Example 5, parsed by the high-speed interrupt channel inside the PLC (response delay < 5ms), and superimposed on the PID output voltage of the basic mode to form a composite speed control signal. The pre-speed-up command represents the relative speed increment as a percentage (e.g., +18%), corresponding to a linear increase in the current operating speed of the EC fan. The activation condition for this mode is that the LSTM model determines the winding temperature rise ΔT within the next 60 seconds. 60 The temperature will exceed the threshold of 1.2K, and the temperature difference ΔT between the inlet and outlet of the cooling water has shown a continuous upward trend for 2 seconds (slope > 0.15K / s). Its functional positioning is to break the waiting-response inertia of traditional temperature control and advance the fan action to the initial stage before the temperature rises significantly. The linkage relationship is as follows: the LSTM model identifies the load change characteristics based on the primary side current harmonic distortion rate THD and the active power change rate dP / dt, and predicts the thermal inertial response by combining the winding temperature rise history curve, and outputs the pre-speed command. The EC fan increases the air volume in advance accordingly, and strengthens the heat exchange end capacity before the cooling water temperature rises, so as to form a water-air dual-channel feedforward synergy with the ΔQ flow regulation on the cooling plate 5 side, and jointly suppress the temperature rise peak. The working result is that the EC fan starts to speed up 27 seconds before the return water temperature actually rises, which shortens the overall thermal response time of the heat exchange unit 6 from 42s to 15s.

[0082] The natural cooling intervention mode is as follows: When the ambient wet-bulb temperature sensor (model: VAISALAHMP155, accuracy ±0.2℃@25℃) detects that the ambient wet-bulb temperature is lower than the set cooling water temperature (default 30℃), and the temperature difference ΔT between the cooling water inlet and outlet is continuously greater than 3K (sampling period 1s, duration ≥5s), the PLC automatically performs the following actions: ① Outputs a shutdown signal to the circulating pump 7 driver to cut off the cooling water circulation; ② Switches the EC fan to constant airflow mode, maintaining the airflow at 35%~45% of the rated value (dynamically adjusted according to the magnitude of ΔT); ③ Closes the bypass valve of the plate heat exchanger 8, causing the cooling water to remain stagnant in the heat exchange unit 6; at this time, the system enters air-water automatic cooling mode. In the convective heat transfer mode, the fan is only used to drive air across the surface of the stationary water tank, relying on the temperature difference between water and air to drive heat transfer. The physical basis of this mode is that when the wet-bulb temperature is low enough, the maximum possible heat absorbed by the air (i.e., enthalpy difference) is sufficient to remove the current heat dissipation of the transformer, without the need for a water pump to drive water circulation. Seamless switching between this mode and the basic mode and the feedforward enhancement mode is achieved through the PLC's internal state machine. The switching process is undisturbed and does not trigger alarms. The results show that under typical spring and autumn operating conditions (ambient wet-bulb temperature 12-18℃), this mode is used for an average of 6.2 hours per day, reducing the system's average daily comprehensive energy consumption by 12%. Furthermore, due to the elimination of water circulation, the risk of mechanical wear and leakage is reduced simultaneously.

[0083] Specifically, the three speed control modes are uniformly managed by the PLC's built-in multi-state priority scheduling module: the basic mode is always enabled as a fallback control; the feedforward enhancement mode is automatically superimposed when the LSTM trigger condition is met, and its instruction weight is higher than that of the basic mode, but it does not cover the integral anti-saturation mechanism of the PID; the natural cooling intervention mode has the highest priority, and once activated, it forcibly blocks the other two modes and locks the circulation pump 7 to the off state; all mode switching is completed based on hard-wired signals and millisecond-level scan cycles, with no software delay; the operating parameters of each mode (such as PID tuning values, LSTM confidence thresholds, and wet-bulb temperature compensation offsets) are all stored in the PLC's non-volatile memory area, supporting remote online modification and version backup via the HMI interface.

[0084] As an optional embodiment, the specific implementation of the solution in this application is as follows: At a 110kV substation site in a northern region, the ambient temperature is -5℃, the relative humidity is 78% (corresponding to a wet-bulb temperature of approximately -7.2℃), the system is operating in steady state at 40% load, the cooling water is set at 30℃, and the measured ΔT=4.1K. At this point, the PLC detected that the ambient wet-bulb temperature (-7.2℃) was <30℃ and ΔT >3K, and immediately activated the natural cooling intervention mode: the circulating pump 7 stopped, the EC fan speed decreased from 2800rpm to 1450rpm (corresponding to 38% of the rated air volume), and the cooling water remained stagnant in the heat exchange unit 6; infrared thermal imaging showed that the average surface temperature of the heat exchange unit 6 slowly rose from 32.6℃ to 34.3℃ and then stabilized, with ΔT maintained in the range of 4.0~4.2K; after running continuously for 3 hours, the system did not issue an over-temperature alarm, and the winding fiber optic temperature measurement showed that the temperature rise fluctuation was ≤0.3K; when the ambient wet-bulb temperature rose back to above 28℃, the PLC automatically exited the mode, restarted the circulating pump 7, and resumed the basic closed-loop regulation mode. This process was completed without human intervention, verifying the reliable start-up and shutdown and thermal balance maintenance capabilities of the natural cooling intervention mode under low temperature and high humidity weather conditions.

[0085] Through the above technical solutions, this application achieves the following: Since the EC fan has high-resolution speed regulation capability and dual protocol compatibility, the basic mode can realize fine closed-loop control of cooling water return temperature; Since the feedforward enhancement mode receives the pre-speed-up command output by the LSTM model and increases the fan speed before the return water temperature rises, it effectively compensates for the control lag caused by the thermal inertia of the cooling medium, making the fan response 27s earlier, forming a water-air dual-channel coordinated acceleration with the flow regulation on the cooling plate 5 side; Since the natural cooling intervention mode shuts down the circulating pump 7 and only uses the fan for air-water natural convection heat exchange when the ambient wet-bulb temperature is lower than the cooling water set temperature and ΔT>3K, it actively eliminates water pump energy consumption under suitable meteorological conditions, reducing the average daily comprehensive energy consumption by 12%, while also reducing the system mechanical failure rate.

[0086] Example 8: In an optional embodiment, this application also provides a direct cooling structure for the transformer body, wherein a spiral coil cooling coil 19 is added inside the transformer oil tank, and its oil inlet and outlet are connected to the cooling water circuit of the closed-loop temperature control system through a plate heat exchanger 8. The spiral coil has a pitch-to-diameter ratio of 3.2:1, and the inner wall of the coil is coated with a nano-scale TiO2 photocatalytic coating to inhibit the growth of microorganisms in the oil passage.

[0087] The spiral coil is installed inside the transformer tank, near the lower oil passage area of ​​the high-voltage winding. Its axis is arranged vertically along the height of the tank, forming a single-headed right-handed continuous spiral structure. This spiral coil can be bent from seamless copper-nickel alloy (CuNi70 / 30) tubing, with an outer diameter of φ16mm, a wall thickness of 1.2mm, a total unfolded length of 18.4m, and an effective heat exchange surface area of ​​0.93m². 2 Alternatively, it can be made of 316L stainless steel thin-walled tubing (outer diameter φ14mm, wall thickness 0.8mm) precision-formed by a CNC spring winding machine, with the surface treated by electrolytic polishing (Ra≤0.2μm) to reduce oil flow resistance and improve the uniformity of the photocatalytic activity interface; or it can be made of titanium alloy TA2 tubing (outer diameter φ18mm, wall thickness 1.0mm), suitable for highly corrosive mineral insulating oil environments. The spiral coil has flanged oil inlet and outlet ports welded to both ends, with a nominal diameter of DN25 and a sealing surface type of RF (convex face). It is equipped with flexible metal spiral wound gaskets (Inconel 625 + flexible graphite) to ensure no leakage under all operating conditions from -30℃ to 95℃.

[0088] The pitch-to-diameter ratio of the spiral coil is 3.2:1, which can refer to the ratio of the vertical distance between the central axes of two adjacent coils (i.e., pitch P) to the outer diameter D of the coil being constant at 3.2. When using a tube with an outer diameter D=16mm, the corresponding pitch P=51.2mm. This ratio is not a fixed geometric constraint, but rather the optimal value obtained through ANSYS Fluent multiphase flow-heat transfer coupled simulation and bench test calibration. Its functional positioning is to increase the Reynolds number Re of the oil flow to the range of 4200-5800 while maintaining an oil-side pressure drop increment ≤12kPa (relative to the state without coil), triggering a fully developed turbulent heat transfer state, while suppressing local stagnation point overheating caused by secondary vortex shedding. Under this ratio, the helix angle α≈11.3° ensures that the oil flow forms a stable axial + circumferential composite disturbance along the tube wall, while avoiding flow separation and bubble accumulation caused by an excessively large helix angle.

[0089] The oil inlet and outlet of the spiral coil are connected to the cooling water circuit of the closed-loop temperature control system via a plate heat exchanger 8. The plate heat exchanger 8 is a brazed all-stainless steel structure (316L corrugated plates + 316L frame), with a rated heat exchange capacity of 45kW, a logarithmic mean temperature difference (LMTD) of 8.2K, a design pressure of 1.6MPa on the hot side (oil circuit), and a design pressure of 1.0MPa on the cold side (water circuit). The inlet and outlet of the oil circuit are respectively connected to the upper oil return chamber and the lower oil inlet chamber of the transformer tank, forming a counter-current heat exchange path from top to bottom. The water circuit is connected in parallel to the main water supply branch and the return water branch of the closed-loop temperature control system. The connection point is located downstream of the outlet of the circulating pump 7 and upstream of the expansion tank 14, ensuring that the cooling water flow and temperature are uniformly regulated by the intelligent control unit 9. This connection method allows the spiral coil to not change the original natural convection main path of the oil circuit, but only to intervene as a bypass reinforcement unit to enhance heat exchange. The system has strong compatibility and does not require modification of the transformer body structure.

[0090] The inner wall of the spiral coil is coated with a nanoscale TiO2 photocatalytic coating. This coating can be anatase TiO2 nanoparticle film grown in situ using the sol-gel method, with a thickness of 80–120 nm and a grain size of 12–18 nm. XRD and TEM confirmed that it has a high specific surface area (≥95 nm). 2 / g) and strong ultraviolet response capability; it can also be a TiO2 / TiN stacked structure prepared by atomic layer deposition (ALD) process, with a bottom TiN layer 5nm thick to enhance interfacial bonding and an upper TiO2 layer 60nm thick, which also has the function of extending visible light response; or a nitrogen-doped TiO2 micro-nano composite coating prepared by plasma spraying method, with a nitrogen content of 1.8at%, and the band gap width reduced from 3.2eV to 2.7eV, which can withstand the trace ultraviolet radiation (λ=320~380nm, intensity about 0.15mW / cm) leaked during transformer operation. 2 The coating continuously generates active oxygen (·OH, H2O2) under the shearing action of oil flow, achieving an inhibition rate of over 99.2% against common sulfate-reducing bacteria (SRB), iron bacteria (IB), and saprophytic bacteria (TGB) in oil (tested according to ISO11737-1:2018). The coating does not change the flow cross section inside the pipe and has been verified by the ASTM D130 copper strip corrosion test to have no catalytic aging effect on mineral insulating oil.

[0091] The spiral coil works in conjunction with the closed-loop temperature control system during operation: when the intelligent control unit 9 detects that the transformer load rate is continuously higher than 85% and the temperature rise gradient at the top of the distributed optical fiber temperature field of the winding is greater than 1.2℃ / cm, it automatically increases the speed of the circulating pump 7 and simultaneously adjusts the cold side water flow of the plate heat exchanger 8 to maintain the cooling water inlet temperature at 22±1℃; at this time, the high-temperature oil enters the coil from the top of the oil tank, and forms a spiral-axial composite flow state under the dual action of centrifugal force and boundary layer disturbance in the spiral channel, which enhances the heat conduction between the oil and the pipe wall; the heat is conducted through the pipe wall to the cold side of the plate heat exchanger 8 and carried away by the low-temperature cooling water; at the same time, the oil flow continuously washes the surface of the TiO2 coating, and the trace ultraviolet components in the oil stimulate photocatalytic reaction, inhibiting the attachment of microorganisms and the formation of biofilm precursors, ensuring the long-term stable heat exchange performance of the coil.

[0092] As an optional embodiment, the specific implementation of this application is as follows: Inside the oil tank of an S13-M-630 / 10 oil-immersed transformer, a CuNi70 / 30 spiral coil with an outer diameter of φ16mm and a wall thickness of 1.2mm is installed in the oil passage space below the high-voltage winding (approximately 1000mm × 400mm × 350mm). The coil has a total of 36 turns and a pitch of 51.2mm. The coil axis is 280mm above the bottom plate of the oil tank. The coil's inlet and outlet are led out of the oil tank via DN25 stainless steel short pipes and connected to a brazed plate heat exchanger of model APVM30-BR. Heat exchanger 8; The cold side of the heat exchanger is connected in parallel to the main water circuit of the closed circulation system through a φ40 stainless steel pipe, with the connection point 1.2m away from the outlet of the circulation pump 7; The inner wall of the coil is coated with anatase TiO2 nano-coating by sol-gel method and cured by heat treatment at 300℃ for 2h; After the system is put into operation, under the impact condition of 120% rated load for 2s, the temperature rise of the top oil layer drops from 89.4℃ without the coil to 87.6℃. With the synergistic effect of the external cooling plate 5, the peak temperature rise measured by the high voltage winding FBG is further reduced by 3.2℃, which verifies the effectiveness of the internal and external dual-path cooling.

[0093] Through the above technical solution, this application achieves the following: a spiral coil cooling coil 19 linked to a closed-loop temperature control system is added inside the transformer oil tank, and its pitch-to-diameter ratio is limited to 3.2:1, which enhances the turbulent heat transfer intensity inside the oil circuit and improves the heat transfer efficiency without significantly increasing the system pressure loss; a nano-level TiO2 photocatalytic coating is set on the inner wall of the coil, which continuously generates active oxygen under the action of oil flow shear and trace ultraviolet radiation, inhibiting the growth of microorganisms and avoiding heat transfer attenuation caused by biofilm blockage; thus solving the technical problem that it is difficult to fully eliminate the internal temperature gradient and local hot spots of the oil circuit by relying solely on the external cooling plate 5, and achieving the organic integration and long-term stable operation of the transformer oil-water-air three-stage heat dissipation path.

[0094] Example 9: In another optional embodiment, this application also provides that the human-machine interface of the intelligent control unit 9 is configured with a three-dimensional thermal cloud map display module, the data source of which includes: a distributed optical fiber temperature field on the surface of the cooling plate 5; a transformer winding fiber optic grating temperature sensor array 20; and pressure and flow sensors 10 at the inlet and outlet of each cooling chamber. The 3D thermal cloud map supports timeline backtracking, automatic hotspot labeling, and early warning of cooling efficiency decay trends.

[0095] The 3D thermal cloud map display module is a core functional unit in the human-machine interface of the intelligent control unit 9, used to realize the spatial visualization of multi-source temperature and fluid parameters. Its function is to map discrete, heterogeneous, and non-spatial aligned sensor data to a unified 3D coordinate system to form a dynamic thermal distribution image with physical location correspondence, thereby supporting maintenance personnel to carry out spatial positioning diagnosis and trend status assessment. This module does not constitute an independent hardware device, but is deployed in software form in the human-machine interface device integrated in the intelligent control unit 9, and interacts with the PLC edge controller in real time via industrial Ethernet (PROFINET protocol).

[0096] The distributed fiber optic temperature field on the surface of cooling plate 5 refers to the two-dimensional temperature distribution dataset collected by distributed fiber optic temperature sensors 17 arranged along the surface of cooling plate 5 with a spatial resolution of 10 cm. Its measurement accuracy is ±0.5℃, and the data sampling period is 2s. This temperature field is mapped in the three-dimensional thermal cloud map as the temperature color scale distribution on the plane (XY plane) where cooling plate 5 is located. The height in the Z direction is constant, and it only represents the temperature amplitude. This data is used to reflect the near-field heat transfer uniformity between cooling plate 5 and transformer winding. Its spatial continuity and resolution directly determine the reliability of local hot spot identification in the thermal cloud map. In this embodiment, the fiber optic sensor is attached to the surface of the metal substrate on the leeward side of cooling plate 5 and fixed with high-temperature resistant epoxy glue (Tg=150℃) to avoid signal drift caused by thermal expansion displacement. Its layout path strictly corresponds to the division of cooling chambers, and each cooling chamber covers no less than 6 measuring points to ensure the representativeness of the regional temperature. This temperature field data can be synchronized with other data sources with timestamps (error ≤10ms) to support multi-source fusion interpolation.

[0097] The transformer winding fiber grating temperature sensor array 20 can refer to a discrete array of fiber grating (FBG) sensors embedded in the insulation layer at the end of the high-voltage winding, with a total of 16 measuring points, staggered along the winding axis and circumference, a spatial resolution of 10 cm, and a temperature measurement accuracy of ±0.3℃. This array is mapped to the corresponding spatial coordinate position of the transformer winding geometric model in the three-dimensional thermal cloud map, forming a temperature node cloud in the YZ or XZ section. Its function is to provide the true temperature rise response of key hot spots inside the winding body, making up for the limitation that the surface temperature of the cooling plate 5 cannot reflect the deep thermal inertia of the winding. Each FBG sensor is connected in series to a demodulator (sampling rate of 10 Hz) through a single-mode fiber, with a wavelength demodulation accuracy of ±1 pm, and is converted into an absolute temperature value through a temperature calibration coefficient. The data of this array and the temperature field of the cooling plate 5 are superimposed and rendered in the thermal cloud map using different color systems (such as cold colors representing the cooling plate 5 and warm colors representing the winding) to distinguish the temperature gradient jump regions on the heat transfer path.

[0098] The inlet and outlet pressure and flow sensors 10 for each cooling chamber can refer to differential pressure flow sensors 10 (accuracy ±1.0%FS) and absolute pressure sensors 21 (range 0~1.0MPa, accuracy ±0.25%FS) installed on the inlet and outlet branches of each cooling chamber, totaling 8 sets (corresponding to 8 cooling chambers). Their output signals are acquired by an analog module (SM1231, 16-bit resolution) and then sent to the PLC. This set of data does not directly participate in the temperature color rendering in the 3D thermal cloud map, but is used as an auxiliary dimension in the thermal efficiency calculation: through the measured flow rate Q of each cooling chamber. With the pressure difference between inlet and outlet ΔP Based on the known flow channel structure parameters, the current heat transfer coefficient h of the cooling chamber can be inverted. The cooling efficiency attenuation index is generated by comparing it with the initial calibration value. This index is mapped to the transparency attenuation or border flicker frequency of the corresponding cooling chamber area in the thermal cloud map, realizing the triple coupling visualization of temperature, fluid and efficiency.

[0099] The construction and operation of the 3D thermal cloud map are based on the OpenGLES 3.0 graphics engine and are executed in real time on the Kunlun Tongtai TPC1061Ti human-machine interface device. The origin of its coordinate system is set at the lower left front face of the cooling plate 5, with the X-axis along the horizontal direction (width direction) of the cooling plate 5, the Y-axis along the vertical direction (height direction), and the Z-axis perpendicular to the cooling plate 5 and pointing towards the transformer body. All sensor data are aligned according to a unified timestamp and then input into an inverse distance weighted interpolation algorithm (IDW, power exponent p=2) to generate a regular grid temperature matrix with a spatial resolution of 10cm×10cm. During the interpolation process, the fiber optic data of the cooling plate 5 is assigned a basic weight of 1.0, and the winding FBG data is weighted by a distance attenuation weight based on its spatial projection distance (weight = 1 / (1+d)). 2), where d is the projected Euclidean distance in cm), and the pressure and flow data are only used for post-processing calculations and do not participate in the interpolation weight allocation; after interpolation, the temperature matrix is ​​linearly mapped to the RGB color space (0℃→dark blue, 70℃→bright red), and semi-transparent isotherms (5℃ intervals) and vector streamlines (synthesized based on the flow direction of each cooling chamber) are superimposed.

[0100] As an optional embodiment, the specific implementation of this application is as follows: After the system is powered on, the PLC collects multi-source sensor data every 2 seconds and pushes it to the human-machine interface via the PROFINET protocol; after receiving the data, the interface engine first completes the timestamp alignment and outlier removal (using the 3σ criterion), and then starts IDW interpolation calculation to generate the current thermal cloud image frame; when the temperature of a certain cooling chamber area is higher than 108% of the average value of its 8 adjacent grids for 10 consecutive seconds, and the absolute temperature is >65℃, the hot spot automatic marking logic is triggered: a red flashing mark (frequency 2Hz) is superimposed at the center of the area, and a pop-up window displays: #3 cooling chamber hot spot, temperature 72.4℃, it is recommended to check the installation status of the flow guiding component and the working status of the corresponding venturi valve; at the same time, the system continuously records the temperature difference ΔT between the inlet and outlet of each cooling chamber. The hourly average, with its initial 72-hour stable operation phase ΔT The mean is the baseline value ΔT0 When a certain cooling chamber continuously for 24 hours, ΔT Mean compared to ΔT0 When the temperature rises by ≥1.5K, a warning of cooling performance degradation trend will be activated: a yellow exclamation mark icon will be displayed in the upper right corner of the thermal cloud map, and a maintenance reminder will be pushed: the contamination level of area 5#3 of the cooling plate is under assessment, and cleaning and maintenance are expected to be required. It is recommended to perform the maintenance within 72 hours.

[0101] Through the above technical solutions, this application achieves the following: The three-dimensional thermal cloud map integrates three heterogeneous data sources: the surface temperature field of the cooling plate 5, the FBG array inside the winding, and the fluid parameters of each cooling chamber. It performs high-resolution interpolation and multi-dimensional attribute mapping in a unified three-dimensional space, which can intuitively present the complete heat transfer path and bottleneck links between the cooling system and the transformer body; The hot spot automatic labeling algorithm is based on the dual criteria of neighborhood statistics and time persistence, which can eliminate transient interference and accurately locate the real thermal anomaly location; The cooling efficiency decay warning is based on the measured ΔT drift trend, which can issue a warning in the early stage of performance degradation (when the contamination level increases by about 30%), turning passive maintenance into pre-intervention.

[0102] Example 10: In another optional embodiment, the system meets the following performance indicators when operating continuously within an ambient temperature range of -30℃ to 40℃: Under a sudden short-term overload of 120% of the rated load, lasting for 2 seconds, the peak temperature rise of the transformer winding is reduced by ≥25% compared to when LSTM feedforward control is not enabled; b. The actual flow rate deviation of each cooling chamber is ≤ ±3.7%; c. In high humidity environments (RH≥90%), the air intake efficiency retention rate at 25℃ is ≥92%; Winter mode startup time ≤ 5 minutes; The overall energy consumption of e is reduced by ≥30% compared to traditional air-cooled systems, and by ≥22% compared to the original scheme in Appendix 1.

[0103] Among them, index a is a comprehensive reflection of the combined effects of the load feedforward-temperature feedback dual-mode predictive controller, the deployment capability of the lightweight LSTM model, the feedforward enhancement mode of the variable frequency fan 18, and the synergistic effect of the spiral coil cooling coil 19 and the closed-loop system. This index reflects that under sudden load disturbances, the system can detect the current harmonic distortion rate THD and the active power change rate dP / dt in advance, and output the target flow increment ΔQ and the fan pre-speed command before the winding temperature rises significantly, so as to drive the cooling chambers of the cooling plate 5 to distribute the flow as needed and increase the heat exchange intensity, thereby suppressing the local heat accumulation in the winding. Its value ≥25% comes from the measured comparison data (73.6℃ vs 102.3℃) under the 120% impact load condition in Example 2. This result can be reproduced under the same transformer model (S13-M-630 / 10), the same ambient temperature and humidity conditions, and the same sampling frequency (10Hz).

[0104] Indicator b is the result of the combined effect of the modular cooling unit, the axial floating sealing joint, the Venturi-diaphragm differential pressure regulating valve 11, and the golden ratio gradient contraction flow channel. Specifically, the Venturi-diaphragm differential pressure regulating valve 11 automatically adjusts the throttling area using water kinetic energy, keeping the pressure difference between any two adjacent cooling units constantly controlled within the range of 0.78–0.82 kPa. The axial floating sealing joint has a built-in disc spring that allows ±0.5 mm axial displacement, compensating for micro-deformations caused by thermal expansion and contraction. The gradient contraction flow channel optimizes the cross-sectional change rate based on λ=0.618, suppressing boundary layer separation and reducing the local resistance coefficient. This structural combination ensures that the cooling unit operates within a range of 2–12 m... 3 The actual flow deviation of each cooling unit within a wide flow range of / h remained stable within ±3.7%. This value is the average statistical range of the synchronous flow monitoring of the eight cooling units in the bench test of Example 3, with a measurement resolution of 0.02m. 3 / h, which meets the repeatability requirements of ISO 5167 for steady-state flow testing.

[0105] Index c represents the functional characterization of the three-stage wetting gradient condenser plate 12 and the capacitive continuous liquid level sensor 13 working in tandem; the superhydrophobic coating on the air inlet side (contact angle > 160°, roll-off angle < 3°) promotes the rapid aggregation and beading of condensate water; the central laser microtextured hydrophilic zone (micro-pit diameter 20μm, depth 5μm) provides capillary driving force to guide the directional migration of the water film; the graphene-titanium dioxide photocatalytic hydrophobic layer on the air outlet side has both antibacterial and UV self-cleaning functions, preventing organic matter from depositing and clogging; A capacitive liquid level sensor (0.1mm resolution) enables precise drainage at the drop level, avoiding blockage of the air intake channel caused by water film bridging. In a constant humid and hot environment with RH≥90% and 25℃, this structure ensures that the air intake efficiency remains stable at ≥92%. This value is derived from continuous monitoring data for 72 hours in the artificial climate chamber (GB / T2423.3) in Example 4. The air intake efficiency in the initial clean state is used as the benchmark (100%), and the air volume attenuation rate is recorded every 2 hours. The average value is then used to calculate the retention rate.

[0106] Index d is the timeliness verification of the coordinated response of the paraffin-based phase change cold storage module 16 and the micro-heating wire mesh of the magnetically coupled variable frequency pump; the PCM module (phase change temperature 5℃) in the expansion tank 14 stores cold in summer and releases cold in winter, pre-cools the pipeline and inhibits the crystallization of ethylene glycol solution; when the ambient temperature drops to -30℃, the system detects that the return water temperature is ≤-10℃ and the real-time flow rate is >15% lower than the rated value, that is, the viscosity compensation mode is activated, and the liquid layer around the impeller is locally heated by the micro-heating wire, so that the dynamic viscosity of this area is reduced by ≥40%, and the pump efficiency is restored to more than 92%; thus, the entire process of injecting cooling medium, venting, and pressurizing to 0.35MPa from a completely emptied state is achieved in a start-up time of ≤5min; this value is the average time (4 minutes and 18 seconds) of the three repeated low-temperature start-up tests in Example 5, with the timing starting point being the start command issued by the PLC and the ending point being the system stably outputting the rated flow rate and ΔT entering the set bandwidth (±0.5K).

[0107] Indicator e represents the system-level energy-saving effect resulting from the superposition of all technical features; among them, a reduction of ≥30% compared to traditional air-cooled systems refers to the reduction in energy consumption per unit time of this system compared to the forced air-cooling scheme disclosed in CN201720321549.8, under the same load spectrum (including a sudden increase of 50% under no-load conditions, a short-term impact of 120%, and day-night temperature difference cycles) and the same environmental boundary conditions; a reduction of ≥22% compared to the original scheme in Annex 1 refers to the reduction in energy consumption per unit time of this system under the same equipment platform (ZBW-12 / 0.4–6301–ⅠA) and the same testing conditions. Within a 72-hour period, the energy consumption baseline value of this system compared to Appendix 1 (without integrated LSTM feedforward, without venturi valve, without three-section condenser plate 12, and without PCM + micro heating wire) decreased. Both data points are from energy efficiency test reports issued by a third-party metrology institution (CMA certified). The test was conducted based on Appendix B of DL / T572—2021 "Operating Procedures for Power Transformers" and GB / T1019—2008 "Energy Efficiency Limits and Energy Efficiency Grades for Electric Fans", with a measurement accuracy better than ±0.5%.

[0108] Specifically, the five performance indicators mentioned above are not isolated but constitute a mutually supportive and closed-loop verified system of technical achievements: Indicator a verifies the effectiveness and timeliness of feedforward control; Indicator b ensures the engineering feasibility of multi-cooling chamber flow distribution; Indicator c resolves the contradiction between condensate management and air intake efficiency in high humidity environments; Indicator d overcomes the system startup bottleneck under extreme low temperature conditions; and Indicator e confirms the economic advantages of the overall solution from the perspective of energy consumption throughout the entire life cycle. All five point to the same technical essence—through multi-physics coupling modeling, flow channel constitutive optimization, interface wetting control, and phase change-rheological coordinated control, achieving an integrated performance leap in the heat dissipation system under wide temperature range, high humidity, and strong dynamic load scenarios, characterized by fast response, accurate flow distribution, stable drainage, rapid startup, and low energy consumption.

[0109] As an optional embodiment, the specific implementation of this application's solution is as follows: This system is deployed on-site at a high-altitude prefabricated box-type substation (altitude 3200m, average annual temperature 2.8℃, extreme low temperature -30.2℃, extreme high temperature 38.6℃). The average annual humidity is 78%, but the rainy season RH≥90% lasts for 117 days. The system operates continuously for 18 months, during which it experiences 5 120% rated load impacts (each lasting 2s), 12 diurnal temperature difference cycles (-28℃→37℃), and 37 thunderstorm and high humidity conditions (RH≥95%, lasting 4~18h). Operation and maintenance records show that the peak winding temperature rise is always below 85℃ (national standard limit 105℃), the measured average flow deviation of the cooling chamber is ±3.2%, the minimum air intake efficiency is maintained at 92.7%, the start-up time on the coldest day in winter is 4 minutes and 41 seconds, and the average annual comprehensive energy consumption is reduced by 31.4% compared to the original air-cooled system in the station and by 23.8% compared to the solution in Appendix 1. All data is automatically collected by the station-side SCADA system, encrypted and uploaded to the cloud platform by the edge gateway, and digitally signed and stored, ensuring complete traceability.

[0110] Through the above technical solutions, this application achieves the following: Due to the integration of a load-feedforward-temperature-feedback dual-mode predictive controller and its cooperating hardware actuator, the cooling process can be intervened in advance under sudden short-term overload, reducing the peak temperature rise of the winding by ≥25%; Due to the adoption of a Venturi-diaphragm differential pressure regulating valve 11 combined with an axial floating sealing joint, the flow deviation of each cooling chamber can still be maintained at ≤±3.7% during long-term operation of the modular cooling unit; Due to the construction of a hydrophobic-hydrophilic-photocatalytic three-stage wetting gradient condenser plate 12 and matching it with a capacitive continuous liquid level sensor, the air intake efficiency is maintained at ≥92% in a high humidity environment with RH≥90%; Due to the integration of a paraffin-based PCM module in the expansion tank 14 and the embedding of a micro-heating wire in the impeller channel of the circulating pump 7, the winter mode start-up time is ≤5min at an extreme low temperature of -30℃; Due to the system-level coupling application of all the aforementioned energy-saving technologies, the overall energy consumption is reduced by ≥30% compared to the traditional air-cooled system and by ≥22% compared to the original scheme in Appendix 1.

Claims

1. An integrated intelligent temperature control and heat dissipation system for a prefabricated box-type substation, comprising a base 1 and a housing 2 mounted on the base 1, wherein the housing 2 is divided into a low-voltage chamber, a transformer chamber, and a high-voltage chamber by a partition 3; characterized in that, The transformer room is equipped with an integrated cooling unit 4, which is connected to a closed-loop temperature control system; the system also includes an intelligent control unit 9 for coordinating the operation of each subsystem. The closed-loop temperature control system includes: Cooling plate 5 is installed in the transformer room and faces the transformer equipment. The interior is divided into multiple cooling chambers, and air inlets are provided between the cooling chambers. Heat exchange unit 6, located outside the housing 2, includes a closed cooling tower or a dry heat exchanger, and has a built-in variable frequency fan 18; The circulating pump 7 is connected to the outlet of the cooling plate 5 and the inlet of the heat exchange unit 6 to drive the cooling water circulation. Expansion tank 14 and water replenishment device 15; Plate heat exchanger 8, which can be optionally installed in the cooling water circuit; The intelligent control unit 9 is equipped with a load feedforward-temperature feedback dual-mode predictive controller. This controller embeds a lightweight long short-term memory neural network LSTM model. The input variables include the transformer primary current harmonic distortion rate (THD), the active power change rate (dP / dt), the winding temperature rise history curve measured by distributed optical fiber, and the cooling water inlet and outlet temperature difference (ΔT). The output is the target flow increment ΔQ of each cooling chamber in the next 60 seconds and the pre-speed command of the variable frequency fan 18 of the heat exchange unit 6.

2. The system according to claim 1, characterized in that, The cooling plate 5 adopts a modular and detachable structure, which is composed of multiple independent cooling units spliced ​​together in the horizontal direction; each cooling unit corresponds to a cooling chamber, and the units are connected to the water channel through an axial floating sealing joint. The axial floating sealing joint has a built-in disc spring, which allows ±0.5mm axial displacement to compensate for thermal expansion and contraction. Each cooling unit has an integrated Venturi-diaphragm differential pressure regulating valve 11 at its water inlet. The Venturi-diaphragm differential pressure regulating valve 11 uses the kinetic energy of water to drive the deformation of the elastic diaphragm and automatically adjusts the throttling area to keep the pressure drop difference between any two adjacent cooling units constant at ≤0.8kPa. The cross-section of the internal flow channel of each cooling unit shrinks in a gradient according to the golden ratio λ=0.

618.

3. The system according to claim 1 or 2, characterized in that, The air inlet of the cooling plate 5 is provided with a multi-layer protective structure, and the surface of its condenser plate 12 is divided into three functional zones along the airflow direction: The air inlet side 1 / 3 area is coated with a fluorosilane superhydrophobic coating with a contact angle >160° and a roll-off angle <3°; the middle 1 / 3 area is laser microtextured to form a hydrophilic region with a micro-pit array diameter of 20μm and a depth of 5μm; the air outlet side 1 / 3 area is coated with a graphene-titanium dioxide photocatalytic hydrophobic layer. The bottom of the condenser plate 12 is connected to a capacitive continuous liquid level sensor 13 with a measurement resolution of 0.1 mm. Its output signal is linked to a micro drainage pump to achieve precise drainage at the drop level.

4. The system according to claim 1, characterized in that, The cooling medium used in the closed-loop temperature control system is an aqueous solution of ethylene glycol; the expansion tank 14 also integrates a paraffin-based phase change cold storage module 16 with a phase change temperature of 5°C. The circulating pump 7 is a magnetically coupled variable frequency pump with an impeller made of carbon fiber reinforced PEEK. The impeller flow channel has an embedded micro-heating wire mesh, and the power density is 0.8 W / cm³. 2 ; The intelligent control unit 9 is also equipped with a viscosity compensation mode: when the cooling water temperature is detected to be <-10℃ and the real-time flow rate is reduced by >15% compared with the rated value, the micro heating wire is activated to locally heat the 0.5mm liquid layer around the impeller, so that the dynamic viscosity of this area is reduced by ≥40%.

5. The system according to claim 1, characterized in that, The structural parameters of the LSTM model meet the following requirements: number of hidden layer nodes ≤ 64; training dataset is derived from measured time series data under at least three typical working conditions, including: sudden addition of 50% rated load under no-load conditions, short-term impact overload (duration ≤ 3s), and day-night temperature difference cycle (-25℃→40℃); model weight file size < 50KB; deployed in PLC edge controller; and single inference latency < 200ms.

6. The system according to claim 1, characterized in that, The surface of the cooling plate 5 is attached with a distributed optical fiber temperature sensor 17, which has a spatial resolution of 10 cm and a temperature measurement accuracy of ±0.5℃. The data from the optical fiber sensor is fused with wavelet denoising and Kalman filtering and used as the input source for the "winding temperature rise history curve" in the LSTM model.

7. The system according to claim 1, characterized in that, The variable frequency fan 18 of the heat exchange unit 6 is an EC fan, and its speed control strategy includes: basic mode: closed-loop adjustment based on the cooling water return temperature; feedforward enhancement mode: receiving the pre-speed-up command output by the LSTM model and increasing the fan speed before the return water temperature rises; natural cooling intervention mode: when the ambient wet-bulb temperature is lower than the cooling water set temperature and ΔT>3K, the circulation pump 7 is turned off, and only the fan is used for air-water natural convection heat exchange.

8. The system according to claim 1, characterized in that, The direct cooling structure of the transformer body also includes: a spiral coil cooling coil 19 is added inside the transformer oil tank, and its oil inlet and outlet are connected to the cooling water circuit of the closed-loop temperature control system through a plate heat exchanger 8; the ratio of the spiral coil pitch to the tube diameter is 3.2:1, and the inner wall of the tube is provided with a nano-level TiO2 photocatalytic coating to inhibit the growth of microorganisms in the oil circuit.

9. The system according to claim 1, characterized in that, The human-machine interface of the intelligent control unit 9 is equipped with a three-dimensional thermal cloud map display module. Its data sources include: a distributed optical fiber temperature field on the surface of the cooling plate 5; a transformer winding fiber optic grating temperature sensor array 20; and pressure and flow sensors 10 at the inlet and outlet of each cooling chamber. The three-dimensional thermal cloud map supports time-axis backtracking, automatic hot spot labeling, and early warning of cooling efficiency decay trends.

10. The system according to claim 1, characterized in that, When the system operates continuously within an ambient temperature range of -30℃ to 40℃, it meets the following performance indicators: (a) Under sudden short-term overload (120% rated load, lasting 2s), the peak temperature rise of the transformer winding is reduced by ≥25% compared with the case where LSTM feedforward control is not enabled; (b) The actual flow rate deviation of each cooling chamber is ≤ ±3.7%; (c) Air intake efficiency retention rate ≥92% under high humidity environment (RH≥90%, 25℃); (d) Winter mode startup time ≤ 5 minutes; (e) The overall energy consumption is reduced by ≥30% compared to the traditional air-cooled system and by ≥22% compared to the original scheme in Annex 1.

Images