Offshore wind turbine cabin transformer heat dissipation collaborative optimization system

Abstract

The application discloses a kind of offshore wind power cabin transformer heat dissipation collaborative optimization system, it is related to offshore wind power equipment technical field, including following module: multidimensional data acquisition module, generates multidimensional state parameter set;Flow field reconstruction and positioning module, reconstructs cabin temperature field and flow field distribution using derivative enhancement depth operator network, locates heat dissipation dead space coordinate;Temperature rise prediction and grading module, obtains temperature rise prediction value and heat dissipation requirement grade;Collaborative control decision module, improved TD3 algorithm is introduced into safety critical network to evaluate out-of-limit risk and output safety correction factor, and the optimal collaborative control instruction set is solved;Instruction execution module, adjust air duct baffle, fan frequency and phase change material;Safety protection module, monitor temperature and limit load capacity.The application solves the problem that heat dissipation dead zone positioning is not accurate and control lag, improves heat dissipation efficiency and safety.

Description

Technical Field

[0001] This invention relates to the field of offshore wind power equipment technology, and in particular to a coordinated optimization system for heat dissipation of offshore wind turbine nacelle transformers. Background Technology

[0002] As the core hub of power transmission, the operational stability of offshore wind turbine nacelle transformers directly affects the power generation efficiency and economic benefits of wind farms. Because offshore wind farms are typically located in open sea areas far from land, the operating environment is extremely harsh. High humidity, salt spray, and turbulence and vibration necessitate a very compact nacelle design, with transformers often enclosed within narrow spaces, resulting in significantly worse heat dissipation conditions compared to onshore substations. Traditional transformer heat dissipation control methods often rely on a single temperature threshold judgment; when the temperature at a monitoring point exceeds a preset value, the system automatically activates fans or oil pumps for cooling. While simple and reliable, this control method lacks the ability to perceive the complex flow field distribution within the nacelle, often resulting in only passive responses. It fails to identify localized "heat dissipation dead zones" created by structural obstructions or airflow short circuits within the nacelle, leading to uneven temperature distribution of transformer winding hotspots, frequent localized overheating problems, severely accelerating the aging of insulation materials, and shortening equipment lifespan. Furthermore, most existing heat dissipation strategies employ fixed cooling logic, failing to adaptively adjust based on real-time load current changes and dynamic nacelle flow field, easily resulting in wasted heat dissipation resources or delayed cooling.

[0003] With the development of intelligent technologies, some research has begun to explore the use of computational fluid dynamics (CFD) numerical simulation techniques to reconstruct the temperature field of transformers, aiming to achieve precise localization of internal hotspots. However, traditional CFD simulation methods involve complex mesh generation and iterative calculations, resulting in large computational loads and long processing times. This makes it difficult to meet the real-time requirements of heat dissipation control for offshore wind power transformers, and hinders online dynamic monitoring and real-time control. Furthermore, existing intelligent control algorithms, such as traditional reinforcement learning, suffer from Q-value overestimation when dealing with continuous action space problems, easily leading to policy oscillations. Moreover, they often neglect constraints on physical safety boundaries during training, resulting in control actions that risk exceeding limits, making them difficult to directly apply to power equipment control scenarios with extremely high safety requirements. In addition, existing technologies lack effective hardware-software collaborative protection mechanisms when facing sudden temperature rises or extreme operating conditions, often resorting only to emergency shutdown measures, leading to high downtime maintenance costs and making it difficult to balance heat dissipation efficiency and operational safety. Therefore, there is an urgent need for an intelligent system that can reconstruct the flow field in real time, accurately predict the temperature rise trend, and achieve collaborative optimization of multiple heat dissipation methods under safety constraints, so as to solve the shortcomings of existing technologies in dead zone positioning, real-time control and safety protection.

[0004] Therefore, how to provide a heat dissipation optimization system for offshore wind turbine nacelle transformers is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a collaborative optimization system for heat dissipation of offshore wind turbine nacelle transformers. This invention fully integrates key steps such as the construction of a multi-dimensional state parameter set, derivative-enhanced deep operator network flow field reconstruction, temperature rise rate prediction and classification, improved TD3 algorithm decision-making, and safety protection linkage. It constructs an intelligent heat dissipation closed-loop process with precise location of heat dissipation dead zones, multi-physics feature matching, control strategy optimization under safety constraints, and multi-level response. This achieves comprehensive perception, proactive prediction, and collaborative control of the transformer's thermal state under complex and variable offshore conditions. This invention possesses advantages such as strong physical consistency in flow field reconstruction, high accuracy in temperature rise prediction, good adaptability of control strategies, and a robust safety protection mechanism. It can significantly improve the accuracy of heat dissipation dead zone identification, the utilization efficiency of cooling resources, and the system's survivability under extreme conditions. In particular, by improving the TD3 algorithm and introducing a security critique network, a composite loss function is constructed based on the value evaluation of the dual critique network. This effectively solves the policy oscillation problem caused by Q-value overestimation in existing reinforcement learning methods and overcomes the defect of traditional control ignoring physical security boundaries. Thus, it effectively solves the problems of lagging heat dissipation control, difficulty in managing local overheating, and high risk of control actions exceeding limits in existing technologies.

[0006] A heat dissipation optimization system for offshore wind turbine nacelle transformers according to the present invention includes the following modules:

[0007] The multi-dimensional data acquisition module synchronously collects transformer operating status data and environmental status data to obtain a multi-dimensional status parameter set;

[0008] The flow field reconstruction and localization module is used to input a multi-dimensional set of state parameters into a derivative-enhanced deep operator network, extract environmental features and backbone network encoded spatial coordinates, construct derivative-enhanced constraints, reconstruct the temperature field and flow field distribution inside the cabin, and obtain the spatial coordinates of the heat dissipation dead zone.

[0009] The temperature rise prediction and classification module is used to extract the load current data at the current moment, combine the spatial coordinates of the heat dissipation dead zone to perform feature matching of time series and spatial region, calculate the matching degree between heat generation power and heat dissipation resistance, and obtain the predicted value of temperature rise change rate and the corresponding heat dissipation requirement level.

[0010] The collaborative control decision module is used to input the predicted temperature rise rate and heat dissipation demand level into the improved TD3 algorithm. A safety critique network is introduced to assess the risk of exceeding limits on the tentative control actions generated by the Actor network and output a safety correction factor. The composite loss function is constructed by combining the value assessment results of the dual Critic network. The network parameters are updated through gradient backpropagation to solve for the optimal control strategy and obtain the collaborative control instruction set.

[0011] The instruction execution module is used to execute the collaborative control instruction set, adjust the angle of the baffle inside the air duct to change the airflow direction, and adjust the fan power supply frequency and the working state of the phase change material to obtain the transfer result.

[0012] The safety protection module is used to collect real-time winding temperature data after the execution of the coordinated control command set and compare it with a preset threshold. If it exceeds the safe range, the auxiliary protection device will be activated to obtain the limited transformer load capacity value.

[0013] Optionally, the operation of the multidimensional data acquisition module includes:

[0014] Multivariate temperature data, load current data, and inlet air velocity data are collected separately, and moving average filtering is performed on the multivariate temperature data.

[0015] Zero-point drift correction is performed on the load current data, subtracting the static zero-point offset from the current value; outlier removal is performed on the air inlet wind speed data, removing erroneous data points whose values ​​exceed the preset wind speed range.

[0016] The preprocessed data is aligned according to the collection timestamp and merged to generate a multidimensional state parameter set.

[0017] Optionally, the operation of the flow field reconstruction and localization module includes:

[0018] The multidimensional state parameter set is input into the branch network of the derivative-enhanced deep operator network to extract environmental features and map them into a high-dimensional environment encoding vector. At the same time, the high-dimensional position encoding vector is generated by encoding spatial coordinates through the backbone network.

[0019] The dot product of the high-dimensional environment encoding vector and the high-dimensional location encoding vector is calculated. The dot product is then mapped to temperature scalar and airflow velocity scalar through two independent fully connected layers. The temperature scalar and airflow velocity scalar of all spatial coordinate points are then aggregated to generate the temperature field and flow field distribution inside the cabin.

[0020] The temperature gradient magnitude is calculated by applying the temperature scalar to spatial coordinates.

[0021] Search for and mark heat dissipation dead zones in the temperature and flow field distribution inside the cabin;

[0022] Calculate the spatial coordinates of the geometric center of the heat dissipation dead zone, construct derivative-enhanced constraints, calculate the thermal diffusivity, and use the absolute value of the difference between the thermal diffusivity and the preset standard thermal diffusivity as the residual value.

[0023] Determine if the residual value is less than the preset residual threshold. If it is, output the spatial coordinates of the heat dissipation dead zone. Otherwise, update the network parameters in reverse until the residual value meets the condition.

[0024] Optionally, the operation of the temperature rise prediction and grading module includes:

[0025] The load current data at the current moment is extracted from the multidimensional state parameter set, and the heat generation power value at the current moment is calculated.

[0026] The corresponding airflow velocity scalar is extracted based on the spatial coordinates of the heat dissipation dead zone. The difference between the average airflow velocity scalar inside the computer cabin and the airflow velocity scalar inside the heat dissipation dead zone is divided by the average airflow velocity scalar inside the cabin to obtain the heat dissipation drag coefficient.

[0027] The product of the current heat generation power and the heat dissipation resistance coefficient is calculated as the instantaneous heat accumulation factor. The winding hot spot temperature data is extracted from the multidimensional state parameter set, and the slope value that changes with time is calculated as the historical temperature rise trend factor. The summation with the instantaneous heat accumulation factor is used to obtain the predicted value of the temperature rise rate.

[0028] The predicted temperature rise rate is compared with the preset temperature rise rate threshold, and the heat dissipation requirement level is output.

[0029] Optionally, the working process of the collaborative control decision module includes:

[0030] The predicted temperature rise rate of change is concatenated with the heat dissipation demand level and input into the Actor network of the improved TD3 algorithm to calculate the hidden layer input value. The hidden layer feature value is obtained by performing nonlinear transformation. Through three fully connected layers, the output air duct opening adjustment amount, cooling fan speed command and phase change material activation signal are combined to generate tentative control actions.

[0031] The tentative control actions are input into the safety critique network, which calculates and outputs a safety correction factor.

[0032] The one-dimensional state vector and the tentative control action are concatenated into a combined input vector, which is then input into the first and second Critic subnetworks in the dual Critic network. Linear weighted summation and nonlinear activation operations are performed to obtain the value scores of the first and second state actions. The one with the smaller value is selected as the target value evaluation result.

[0033] The composite loss function value is obtained by summing the negative value of the target value assessment result and the safety correction factor. The first derivative of the composite loss function value with respect to the Actor network parameters is calculated as the gradient value. The updated Actor network parameter values ​​are obtained by subtracting the product of the gradient value and the preset learning rate from the Actor network parameter values. The network parameter update step is repeated until the composite loss function value is less than the preset convergence threshold. The tentative control action output by the Actor network is used as the collaborative control instruction set.

[0034] Optionally, the process of inputting the tentative control action into the security critique network, calculating and outputting the security correction factor includes:

[0035] The tentative control actions are input into the safety critique network, and the values ​​of the air duct opening adjustment, cooling fan speed command and phase change material activation signal are extracted to determine whether they exceed the preset safety action threshold range.

[0036] When the value exceeds the upper limit of the preset safety action threshold range, the absolute value of the difference between the value and the upper limit is calculated and multiplied by the preset penalty weight coefficient. When the value is less than the lower limit of the preset safety action threshold range, the absolute value of the difference between the lower limit and the value is calculated and multiplied by the preset penalty weight coefficient. When the value is within the preset safety action threshold range, the penalty item is set to zero. The safety correction factor is obtained by summing the three penalty items corresponding to the air duct opening adjustment amount, the cooling fan speed command, and the phase change material activation signal.

[0037] Optionally, the operation of the instruction execution module includes:

[0038] Read the collaborative control instruction set, calculate the target baffle angle value based on the air duct opening adjustment amount, and drive the baffle to rotate to the target angle value;

[0039] Calculate the target power supply frequency value based on the cooling fan speed command, and adjust the output pulse frequency of the fan driver to the target power supply frequency value.

[0040] When the phase change material activation signal is high, the solenoid valve switch of the phase change material storage device is closed; when the phase change material activation signal is low, the solenoid valve switch is opened.

[0041] The system collects the temperature and flow field distribution inside the cabin after the data collection process, identifies the temperature changes at the spatial coordinates of the heat dissipation dead zone, and determines that the directional cooling effect for the heat dissipation dead zone has been achieved when the temperature changes are greater than the preset cooling threshold. At the same time, the system continuously collects the temperature data of the winding hot spots and calculates the temperature change difference between the temperature data of the winding hot spots at two adjacent collection times. When the temperature change difference is less than or equal to zero, the system determines that the peak heat load transfer has been achieved.

[0042] Optionally, the operation of the security protection module includes:

[0043] Collect real-time winding temperature data after the execution of the collaborative control instruction set, and compare it with the preset maximum temperature threshold. If it exceeds the preset maximum temperature threshold, it is determined to be outside the safe range.

[0044] When the load is determined to be outside the safe range, a voltage reduction command is sent to gradually reduce the transformer tap position, monitor the load current data in real time, calculate the ratio of the load current to the rated current value, and multiply it by the rated capacity of the transformer to obtain the limited transformer load capacity value.

[0045] The beneficial effects of this invention are:

[0046] This invention addresses the problems of uneven flow field distribution and difficulty in locating heat dissipation dead zones within offshore wind turbine nacelle transformers by deploying a multi-dimensional sensor array and constructing a derivative-enhanced deep operator network. It employs branch and backbone networks to extract features and construct derivative-enhanced constraints, enabling rapid reconstruction of the internal temperature and flow fields of the nacelle and generating the spatial coordinates of the heat dissipation dead zones. By extracting load current data and flow velocity characteristics of the heat dissipation dead zones, combined with heat generation power calculation and heat dissipation resistance analysis, it performs temperature rise rate prediction and heat dissipation demand level classification, outputting the predicted temperature rise rate and corresponding level. The prediction results are input into an improved TD3 algorithm, which, combined with a safety critique network, assesses and corrects the risk of exceeding limits for the tentative control actions generated by the Actor network, outputting a set of collaborative control instructions including duct opening, fan speed, and phase change material activation signals. Furthermore, the execution module adjusts the duct baffles, fan frequency, and phase change material state to implement targeted cooling for the heat dissipation dead zones. Finally, a safety protection module collects the real-time winding temperature and compares it with a threshold, executing voltage reduction and capacity limiting operations when the temperature exceeds the safe range. This enables closed-loop optimization of the entire heat dissipation process for offshore wind turbine nacelle transformers, from multi-dimensional perception, precise positioning, intelligent decision-making to safe execution, effectively improving the accuracy of heat dissipation dead zones, the adaptability of cooling strategies, and equipment safety under extreme operating conditions. Attached Figure Description

[0047] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0048] Figure 1 This is a structural diagram of a heat dissipation collaborative optimization system for offshore wind turbine nacelle transformers proposed in this invention;

[0049] Figure 2 This is a flowchart of the flow field reconstruction and heat dissipation dead zone localization based on derivative-enhanced deep operator networks proposed in this invention.

[0050] Figure 3 This is a flowchart of the calculation of collaborative control decision and safety correction factor for the improved TD3 algorithm proposed in this invention. Detailed Implementation

[0051] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0052] refer to Figures 1-3 A heat dissipation optimization system for offshore wind turbine nacelle transformers includes the following modules:

[0053] The multi-dimensional data acquisition module is used to simultaneously collect transformer operating status data and environmental status data by utilizing a multi-dimensional sensor array deployed at transformer winding hot spots, iron core and nacelle air inlet, and to obtain a multi-dimensional status parameter set.

[0054] The flow field reconstruction and localization module is used to input a multi-dimensional set of state parameters into a derivative-enhanced deep operator network, extract environmental features and backbone network encoded spatial coordinates, construct derivative-enhanced constraints, reconstruct the temperature field and flow field distribution inside the cabin, and locate and identify the spatial coordinates of the heat dissipation dead zone based on the maximum value of the temperature gradient.

[0055] The temperature rise prediction and classification module is used to extract the load current data at the current moment, combine the spatial coordinates of the heat dissipation dead zone to perform feature matching of time series and spatial region, calculate the matching degree between heat generation power and heat dissipation resistance, and obtain the predicted value of temperature rise change rate and the corresponding heat dissipation requirement level.

[0056] The collaborative control decision module is used to input the predicted temperature rise rate and heat dissipation demand level into the improved TD3 algorithm. A safety critique network is introduced to assess the risk of exceeding limits of the tentative control actions generated by the Actor network and output a safety correction factor. A composite loss function is constructed by combining the value assessment results of the dual Critic network. The network parameters are updated through gradient backpropagation to solve the optimal control strategy and obtain a collaborative control instruction set including the air duct opening adjustment amount, cooling fan speed command and phase change material activation signal.

[0057] The instruction execution module is used to execute the collaborative control instruction set, adjust the angle of the baffle inside the air duct to change the airflow direction, and adjust the fan power supply frequency and the working state of the phase change material to obtain the directional cooling effect for the heat dissipation dead zone and the transfer result of peak heat load.

[0058] The safety protection module is used to collect real-time winding temperature data after the execution of the coordinated control command set and compare it with a preset threshold. If it exceeds the safe range, the auxiliary protection device will be activated to obtain the limited transformer load capacity value.

[0059] This implementation significantly improves the accuracy and operational safety of transformer heat dissipation control. By deploying a multi-dimensional sensor array and utilizing a derivative-enhanced deep operator network, rapid reconstruction of the temperature and flow fields inside the nacelle is achieved, accurately locating heat dissipation dead zones that are difficult to identify using traditional methods, thus resolving the risk of localized overheating. Combined with temperature rise prediction and a tiered mechanism, a shift from passive response to proactive prevention is realized. The core lies in introducing an improved TD3 algorithm and a safety critique network. When solving for the optimal control strategy, the risk of exceeding limits is evaluated in real time, and a correction factor is output, ensuring the safety boundary of control actions under extreme conditions. The instruction execution module coordinates the adjustment of duct baffles, fan frequency, and phase change materials to achieve targeted cooling and heat load transfer for heat dissipation dead zones. The safety protection module, as the last line of defense, performs closed-loop verification of the winding temperature after execution, further enhancing the system's fault tolerance. This approach effectively overcomes the shortcomings of traditional threshold control, such as lag and poor nonlinear adaptability, significantly extending the transformer insulation life and ensuring stable equipment operation under complex conditions.

[0060] In this embodiment, the working process of the multidimensional data acquisition module includes:

[0061] Temperature sensors, current transformers, and wind speed sensors deployed in a multi-dimensional sensor array at transformer winding hot spots, iron core, and nacelle air inlet are used to collect multi-dimensional temperature data, load current data, and air inlet wind speed data, respectively. Moving average filtering is performed on the multi-dimensional temperature data to eliminate random noise interference.

[0062] Zero-point drift correction is performed on the load current data by subtracting the static zero-point offset from the current value. Outlier removal is performed on the air inlet wind speed data by removing erroneous data points whose values ​​exceed the preset wind speed range. The preset wind speed range is determined by calculating the average value and standard deviation of historical wind speed data, with the average value minus three times the standard deviation as the lower limit and the average value plus three times the standard deviation as the upper limit.

[0063] The preprocessed data is aligned according to the collection timestamp and merged to generate a multidimensional state parameter set.

[0064] In this embodiment, the working process of the flow field reconstruction and positioning module includes:

[0065] The multidimensional state parameter set is input into the branch network of the derivative-enhanced deep operator network to extract environmental features and map them into a high-dimensional environment encoding vector. At the same time, the high-dimensional position encoding vector is generated by encoding spatial coordinates through the backbone network.

[0066] The dot product of the high-dimensional environment encoding vector and the high-dimensional location encoding vector is calculated. The dot product is then mapped to temperature scalar and airflow velocity scalar through two independent fully connected layers. The temperature scalar and airflow velocity scalar of all spatial coordinate points are then aggregated to generate the temperature field and flow field distribution inside the cabin.

[0067] The temperature gradient magnitude is obtained by calculating the numerical derivatives of the temperature scalar in the horizontal and vertical directions with respect to spatial coordinates, and then performing the sum of squares and square roots.

[0068] In the temperature field and flow field distribution inside the cabin, search for regions where the temperature gradient magnitude is greater than a preset gradient threshold and the airflow velocity scalar is less than a preset flow velocity threshold. Regions that simultaneously meet both conditions are marked as heat dissipation dead zones. The preset gradient threshold is calculated by using the average and standard deviation of the temperature gradient distribution data under historical normal operating conditions inside the cabin. The average value plus three times the standard deviation is used as the threshold. The preset flow velocity threshold is calculated by using the average and standard deviation of the airflow velocity distribution data under historical normal operating conditions inside the cabin. The average value minus three times the standard deviation is used as the threshold.

[0069] Calculate the spatial coordinates of the geometric center of the heat dissipation dead zone, construct derivative-enhanced constraints, calculate the sum of the horizontal and vertical second-order partial derivatives of the temperature scalar with respect to the spatial coordinates, divide the sum by the first-order derivative of the temperature scalar with respect to time to obtain the thermal diffusivity, and use the absolute value of the difference between the thermal diffusivity and the preset standard thermal diffusivity as the residual value. The preset standard thermal diffusivity is obtained by consulting the thermophysical property manual of transformer insulation materials to obtain the thermal conductivity and volumetric heat capacity of the material, and then calculating it by dividing the thermal conductivity by the volumetric heat capacity.

[0070] Determine whether the residual value is less than the preset residual threshold. If it is, output the spatial coordinates of the heat dissipation dead zone. Otherwise, use the residual value to calculate the gradient and update the network parameters in reverse until the residual value meets the condition. The preset residual threshold is determined by calculating the reciprocal of the mean square error between the sampled temperature measurement value and the network prediction value.

[0071] This implementation innovatively employs a derivative-enhanced deep operator network for flow field reconstruction, demonstrating significant advantages over traditional computational fluid dynamics numerical simulations and conventional convolutional neural networks. While traditional CFM methods offer high accuracy, their computational time is lengthy, making it difficult to meet the real-time requirements of transformer heat dissipation control. Furthermore, conventional CNNs, when processing continuous physical fields, lack intrinsic constraints on the physical mechanisms, easily leading to blurred boundaries or distorted predictions.

[0072] This invention extracts environmental and spatial coordinate features through branch and backbone networks, respectively, and rapidly generates the temperature and flow fields of the entire cabin using dot product mapping, achieving millisecond-level real-time reconstruction. A specially introduced derivative enhancement constraint mechanism, through calculating the thermal diffusivity residual for gradient backpropagation and parameter updates, forces the network prediction results to conform to the physical laws of heat conduction. This deep integration of physical information and deep learning not only significantly improves prediction accuracy but also endows the model with strong robustness under conditions of few samples or noise interference. Combining the dynamic gradient and flow velocity threshold generated by statistical learning, it can accurately identify heat dissipation dead zones, effectively overcoming the limitations of traditional methods that rely solely on a single temperature threshold to determine local overheating, and providing a reliable spatiotemporal positioning basis for subsequent precise heat dissipation control.

[0073] In this embodiment, the working process of the temperature rise prediction and grading module includes:

[0074] The load current data at the current moment is extracted from the multidimensional state parameter set, and the product of the square value of the load current data and the DC resistance value of the transformer winding is calculated to obtain the heat generation power value at the current moment.

[0075] The corresponding airflow velocity scalar is extracted based on the spatial coordinates of the heat dissipation dead zone. The difference between the average airflow velocity scalar inside the computer cabin and the airflow velocity scalar inside the heat dissipation dead zone is divided by the average airflow velocity scalar inside the cabin to obtain the heat dissipation drag coefficient.

[0076] The product of the current heat generation power and the heat dissipation resistance coefficient is used as the instantaneous heat accumulation factor. The winding hot spot temperature data within the historical preset time window is extracted from the multidimensional state parameter set. The slope value of the winding hot spot temperature data changing with time is calculated as the historical temperature rise trend factor. The instantaneous heat accumulation factor and the historical temperature rise trend factor are summed to obtain the predicted value of the temperature rise change rate. The historical preset time window is set to ten minutes.

[0077] The predicted temperature rise rate is compared with a preset temperature rise rate threshold. When the predicted temperature rise rate is less than the first preset temperature rise threshold, a level 1 heat dissipation demand level is output. When the predicted temperature rise rate is greater than or equal to the first preset temperature rise threshold and less than the second preset temperature rise threshold, a level 2 heat dissipation demand level is output. When the predicted temperature rise rate is greater than or equal to the second preset temperature rise threshold, a level 3 heat dissipation demand level is output. The preset temperature rise rate threshold is determined by statistically analyzing the distribution range of historical temperature rise rate data and dividing it evenly.

[0078] In this embodiment, the working process of the collaborative control decision module includes:

[0079] The Actor network of the improved TD3 algorithm is constructed by concatenating the predicted temperature rise rate of change with the heat dissipation demand level into a one-dimensional state vector input. The one-dimensional state vector is linearly weighted and summed to obtain the hidden layer input value. The hidden layer input value is then transformed by a nonlinear activation function to obtain the hidden layer feature value. The above linear weighting and nonlinear activation operations are iteratively executed through three fully connected layers to output the air duct opening adjustment amount, cooling fan speed command and phase change material activation signal, and combine them to generate tentative control actions.

[0080] The tentative control actions are input into the safety critique network, which calculates and outputs a safety correction factor.

[0081] A one-dimensional state vector is concatenated with an exploratory control action to form a combined input vector, which is then input into the first and second Critic subnetworks of the dual Critic network. In the first and second Critic subnetworks, the combined input vector is subjected to linear weighted summation and nonlinear activation operation to obtain the first state action value score and the second state action value score. The numerical values ​​of the first state action value score and the second state action value score are compared, and the smaller value is selected as the target value evaluation result.

[0082] The composite loss function value is obtained by summing the negative value of the target value assessment result and the safety correction factor. The first derivative of the composite loss function value with respect to the Actor network parameters is calculated as the gradient value. The updated Actor network parameter values ​​are obtained by subtracting the product of the gradient value and the preset learning rate from the Actor network parameter values. The network parameter update step is repeated until the composite loss function value is less than the preset convergence threshold. The tentative control action output by the Actor network is used as the cooperative control instruction set. The preset learning rate is set to 0.001 and the preset convergence threshold is set to 0.01.

[0083] In this embodiment, tentative control actions are input into the safety critique network, which calculates and outputs a safety correction factor. The process includes:

[0084] The tentative control actions are input into the safety critique network, and the values ​​of the air duct opening adjustment, cooling fan speed command and phase change material activation signal in the tentative control actions are extracted respectively. Each value is judged to see if it exceeds the preset safety action threshold range. The preset safety action threshold range is determined by querying the physical limit value range of air duct opening, fan speed and phase change material activation signal in the transformer cooling equipment manual.

[0085] When the value is greater than the upper limit of the preset safety action threshold range, the absolute value of the difference between the value and the upper limit is calculated and multiplied by the preset penalty weight coefficient. When the value is less than the lower limit of the preset safety action threshold range, the absolute value of the difference between the lower limit and the value is calculated and multiplied by the preset penalty weight coefficient. When the value is within the preset safety action threshold range, the penalty item is set to zero. The three penalty items corresponding to the air duct opening adjustment amount, the cooling fan speed command, and the phase change material activation signal are summed to obtain the safety correction factor. The preset penalty weight coefficient is set to 0.1.

[0086] This implementation achieves precise coordinated control and safety constraints for transformer heat dissipation by combining an improved TD3 algorithm with a safety critique network. Temperature rise prediction and demand levels are used to construct a state vector, which is then input into the Actor network to generate tentative control actions involving air ducts, fans, and phase change materials. A safety critique network is introduced to evaluate these actions based on physical limit thresholds and output a safety correction factor. A dual-Critic network takes the minimum value assessment and constructs a composite loss function using the correction factor, optimizing network parameters through gradient backpropagation. This mechanism effectively overcomes the oscillation and lag problems of traditional control strategies in nonlinear, high-lag systems. The dual-Critic mechanism suppresses overestimation bias, and the introduction of the safety correction factor ensures that control actions always remain within the physical safety boundary, achieving an optimal balance between heat dissipation efficiency and equipment safety.

[0087] The improved TD3 algorithm of this invention is similar to the original TD3 algorithm in that both retain the core architecture of the dual-delay deterministic policy gradient, namely, the Actor-Critic network structure, which includes one Actor policy network and two Critic value networks. During training, both utilize the soft update mechanism of the target network to maintain network stability and explore the environment state space by introducing policy noise. Furthermore, both use the minimum value of the two Critic networks to estimate the value, in order to suppress overestimation bias.

[0088] The difference lies in that this invention breaks away from the original TD3 algorithm's limitation of solely evaluating action value through the Critic network. It introduces a safety critique network into the policy update stage of the Actor network, constructing a composite optimization mechanism that combines physical constraints and value assessment. In the original algorithm, the Actor network update relies solely on the gradient backpropagation of state-action pairs by the Critic network, lacking explicit constraints on the physical feasibility of the actions themselves. However, this invention, after the Actor network outputs tentative control actions, does not directly calculate the policy gradient. Instead, it pre-implements a safety critique network module, which, based on the physical limit thresholds of the transformer's heat dissipation equipment, performs item-by-item limit checks on the air duct opening, fan speed, and phase change material activation signal, and quantifies and outputs a safety correction factor. In constructing the loss function, this invention innovatively adds this safety correction factor to the negative of the target value assessment result, constructing a composite loss function that includes value gradient information and a physical penalty term. This guides the Actor network parameters to update in a high-value and safe direction, rather than simply pursuing high value.

[0089] Based on the aforementioned improvements, the beneficial effect of this invention lies in the fact that, through the deep coupling of the security critique network and the composite loss function, the improved TD3 algorithm successfully embeds the physical safety boundary of transformer heat dissipation into the decision-making optimization process of reinforcement learning. This design effectively solves the problem that the original TD3 algorithm is prone to generating extreme control commands that violate the physical limits of the equipment during the exploration process, avoiding the risk of equipment damage caused by fan overspeed or mechanical overreach of air duct opening. At the same time, by dynamically adjusting the preset penalty weight coefficient, a balance is achieved between the pursuit of heat dissipation efficiency and the constraints of safe operation, significantly enhancing the robustness and feasibility of the control strategy under complex and variable operating conditions, and ensuring the safety and compliance of the intelligent decision-making results.

[0090] In this embodiment, the operation of the instruction execution module includes:

[0091] Read the collaborative control instruction set, calculate the target baffle angle value based on the air duct opening adjustment amount, send the rotation angle command, drive the baffle to rotate to the target angle value, and guide the cooling airflow to the spatial coordinate position of the heat dissipation dead zone.

[0092] Calculate the corresponding target power supply frequency value based on the cooling fan speed command, adjust the output pulse frequency of the fan driver to the target power supply frequency value, and change the fan motor speed to the speed value required by the cooling fan speed command.

[0093] When the phase change material activation signal is high, the solenoid valve switch of the phase change material storage device is closed, releasing the phase change material so that it flows through the heat dissipation dead zone to absorb heat; when the phase change material activation signal is low, the solenoid valve switch is opened.

[0094] The system collects the temperature and flow field distribution inside the cabin after the data collection process, identifies the temperature changes at the spatial coordinates of the heat dissipation dead zone, and determines that the targeted cooling effect for the heat dissipation dead zone is achieved when the rate of temperature decrease is greater than the preset cooling threshold. At the same time, the system continuously collects the temperature data of the winding hot spots at preset time intervals, calculates the difference between the temperature data of the winding hot spots at two adjacent collection times to obtain the temperature change difference, and determines that the peak heat load transfer result is achieved when the temperature change difference is less than or equal to zero. The preset cooling threshold is calculated by subtracting three times the standard deviation from the average and standard deviation of the temperature decrease rate of the heat dissipation dead zone in historical successful heat dissipation cases. The preset time interval is set to five seconds.

[0095] In this embodiment, the operation of the security protection module includes:

[0096] Collect real-time winding temperature data after the execution of the collaborative control instruction set, compare the real-time winding temperature data with the preset maximum temperature threshold, and determine that it exceeds the safe range when it is greater than the preset maximum temperature threshold. The preset maximum temperature threshold is obtained by querying the heat resistance grade standard of transformer insulation material to obtain the corresponding upper temperature limit value.

[0097] When the load is determined to be outside the safe range, a voltage reduction command is sent to the on-load tap changer of the transformer to gradually reduce the transformer tap position. The load current data is monitored in real time, and the ratio of the current load current data to the rated current value is calculated and multiplied by the rated capacity of the transformer to obtain the limited transformer load capacity value.

[0098] Example 1: To verify the feasibility and practical effectiveness of this invention in the complex operation and maintenance environment of offshore wind power, it was deployed in the intelligent operation and maintenance management platform of a large-scale offshore wind farm in a coastal province. This wind farm has a total installed capacity of 400MW, consisting of 80 offshore wind turbines with a single unit capacity of 5MW, distributed in an area approximately 25 kilometers from the coastline, facing a harsh marine environment characterized by high temperature, high humidity, high salt spray, and frequent strong typhoons. The nacelle transformer, as a key step-up device for wind turbine generators, is located in a confined, enclosed nacelle space with limited heat dissipation. Furthermore, its operation is constrained by factors such as inconvenient maritime transportation and short maintenance windows, placing extremely high demands on its operational reliability. Traditional heat dissipation control methods are mostly based on a single winding temperature threshold for fan start-stop control, lacking a refined perception of the complex flow field distribution inside the nacelle. This makes it difficult to address the heat dissipation dead zones caused by localized hotspots, and the control strategies are rigid, unable to be dynamically adjusted according to real-time operating conditions. This can easily lead to accelerated insulation aging or even burnout accidents in high-temperature, high-load operating scenarios. During the five-month trial operation, the method of this invention utilized a multi-dimensional sensor array deployed in the cabin, including an infrared thermal imager, a multi-point distributed temperature sensor, a hot-wire anemometer, and a current and voltage transmitter, to collect data in real time on winding temperature, ambient temperature and humidity, cooling fan speed, load current, and airflow velocity at key nodes in the cabin. This constructed a holographic data set of the cabin transformer's heat dissipation status, with an average daily data processing volume exceeding 850MB.

[0099] In practical applications, this invention first uses a derivative-enhanced deep operator network to extract features and reconstruct the flow field from the collected multidimensional state parameters. The system inputs the spatial coordinates within the cabin and the environmental feature vectors collected by sensors into the branch and backbone networks. By calculating the dot product of the high-dimensional feature vectors and mapping them, it quickly generates temperature and flow field distribution cloud maps inside the cabin. Addressing heat dissipation dead zones that traditional methods cannot identify, this invention introduces a dual mechanism of temperature gradient magnitude calculation and flow velocity threshold determination. This successfully identifies multiple local heat dissipation dead zones at the top corner of the transformer and the bends of the cooling ducts, and calculates the geometric center coordinates of these dead zones, providing a spatial positioning basis for subsequent precise heat dissipation. Simultaneously, the system constructs derivative-enhanced constraints, and by calculating the residual value of the thermal diffusivity, it achieves adaptive correction of the operator network parameters, ensuring that the flow field reconstruction accuracy remains high even under complex and variable operating conditions. In the decision-making and control stage, this invention abandons the traditional threshold-triggered logic and adopts an improved TD3 algorithm for collaborative control decision-making. To address the large hysteresis and nonlinear thermal dynamics of offshore wind power, the system inputs predicted temperature rise rate, coordinates of the heat dissipation dead zone, and real-time load status into the Actor network. This generates tentative control actions that include adjusting the duct damper opening, distributing fan group speeds, and determining the activation sequence of phase change materials. Specifically, to overcome the shortcomings of traditional reinforcement learning algorithms in safety boundary control, the system introduces a safety critique network. This network performs real-time risk assessment of the actions output by the Actor network, and combined with the minimum Q-value estimation of the dual critique network, outputs a final cooperative control command adjusted by a safety correction factor. This maximizes heat dissipation efficiency while ensuring transformer insulation safety.

[0100] To quantify the practical application effect of this invention, the project team selected Unit 20, which operates under the most severe conditions in the wind farm, as the main experimental subject, and maintained the original traditional temperature control system on the adjacent Unit 19 as a control group. During a five-month comparative test, key operating indicators of the two units under high-temperature and high-load conditions were recorded in detail. Experimental data showed that under extreme conditions with ambient temperatures exceeding 35 degrees Celsius and unit load rates exceeding 90%, the highest average temperature of the transformer winding of Unit 20, using this invention, was controlled at approximately 78.4 degrees Celsius, while the average temperature of the winding of Unit 19, using traditional control, reached as high as 92.6 degrees Celsius, triggering high-temperature alarms multiple times and significantly shortening the remaining lifespan of the insulation material. Regarding the elimination of heat dissipation dead zones, this invention, through a directional cooling strategy, reduced the average temperature of the dead zone by 11.2 degrees Celsius, effectively avoiding the risk of insulation breakdown caused by localized overheating. Furthermore, the improved TD3 algorithm and the introduction of a security critique network reduced the energy consumption of the cooling system by approximately 18.5%, while shortening the response delay of the cooling system from 45 seconds in traditional control to less than 12 seconds, significantly enhancing the system's dynamic adjustment capability. The following is a detailed performance comparison table of the two units under different typical operating conditions:

[0101] Table 1. Comparison of the overall performance of the method of the present invention and traditional control methods.

[0102]

[0103] As can be seen from the detailed comparative data shown in Table 1, the heat dissipation collaborative optimization system based on the improved TD3 algorithm and security critique network proposed in this invention exhibits significant technical advantages under various operating conditions. Especially in the most challenging scenario of "high temperature and high load," traditional control methods, lacking the ability to detect heat dissipation dead zones, saw hotspot temperatures climb to 108.4℃, easily leading to insulating oil cracking and winding deformation. In contrast, this invention, through flow field reconstruction, accurately locates the dead zone and implements directional cooling, successfully suppressing the hotspot temperature below 90℃ without any temperature-over-limit shutdown events. In extreme environments such as "typhoons," the advantages of this invention are further amplified, with a response time of only 10.5 seconds, nearly four times faster than the traditional method's 42 seconds, effectively addressing the challenge of sudden temperature rises brought by typhoons.

[0104] Furthermore, insulation aging rate, a key indicator for measuring the long-term operational reliability of transformers, is reduced to less than 40% of the average aging rate using traditional methods. This means the theoretical service life of the transformer can be extended by more than double. Considering overall energy consumption data, this invention, through an adaptive collaborative control strategy, optimizes the operating curve of the fan group while ensuring effective heat dissipation, achieving an average energy saving of over 15%, thus realizing a dual improvement in safety and economy. This embodiment fully verifies the effectiveness of this invention in solving problems such as difficulty in identifying heat dissipation dead zones, lagging control strategies, and ambiguous safety boundaries in existing technologies, demonstrating broad prospects for application in the offshore wind power sector.

[0105] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A heat dissipation optimization system for offshore wind turbine nacelle transformers, characterized in that, Includes the following modules: The multi-dimensional data acquisition module synchronously collects transformer operating status data and environmental status data to obtain a multi-dimensional status parameter set; The flow field reconstruction and localization module is used to input a multi-dimensional set of state parameters into a derivative-enhanced deep operator network, extract environmental features and backbone network encoded spatial coordinates, construct derivative-enhanced constraints, reconstruct the temperature field and flow field distribution inside the cabin, and obtain the spatial coordinates of the heat dissipation dead zone. The temperature rise prediction and classification module is used to extract the load current data at the current moment, combine the spatial coordinates of the heat dissipation dead zone to perform feature matching of time series and spatial region, calculate the matching degree between heat generation power and heat dissipation resistance, and obtain the predicted value of temperature rise change rate and the corresponding heat dissipation requirement level. The collaborative control decision module is used to input the predicted temperature rise rate and heat dissipation demand level into the improved TD3 algorithm. A safety critique network is introduced to assess the risk of exceeding limits on the tentative control actions generated by the Actor network and output a safety correction factor. The composite loss function is constructed by combining the value assessment results of the dual Critic network. The network parameters are updated through gradient backpropagation to solve for the optimal control strategy and obtain the collaborative control instruction set. The instruction execution module is used to execute the collaborative control instruction set, adjust the angle of the baffle inside the air duct to change the airflow direction, and adjust the fan power supply frequency and the working state of the phase change material to obtain the transfer result. The safety protection module is used to collect real-time winding temperature data after the execution of the coordinated control command set and compare it with a preset threshold. If it exceeds the safe range, the auxiliary protection device will be activated to obtain the limited transformer load capacity value.

2. The offshore wind turbine nacelle transformer heat dissipation collaborative optimization system according to claim 1, characterized in that, The working process of the multidimensional data acquisition module includes: Multivariate temperature data, load current data, and inlet air velocity data are collected separately. Moving average filtering is performed on the multivariate temperature data to eliminate random noise interference. Zero-point drift correction is performed on the load current data, subtracting the static zero-point offset from the current value; outlier removal is performed on the air inlet wind speed data, removing erroneous data points whose values ​​exceed the preset wind speed range. The preprocessed data is aligned according to the collection timestamp and merged to generate a multidimensional state parameter set.

3. The offshore wind turbine nacelle transformer heat dissipation collaborative optimization system according to claim 1, characterized in that, The working process of the flow field reconstruction and localization module includes: The multidimensional state parameter set is input into the branch network of the derivative-enhanced deep operator network to extract environmental features and map them into a high-dimensional environment encoding vector. At the same time, the high-dimensional position encoding vector is generated by encoding spatial coordinates through the backbone network. The dot product of the high-dimensional environment encoding vector and the high-dimensional location encoding vector is calculated. The dot product is then mapped to temperature scalar and airflow velocity scalar through two independent fully connected layers. The temperature scalar and airflow velocity scalar of all spatial coordinate points are then aggregated to generate the temperature field and flow field distribution inside the cabin. The temperature gradient magnitude is calculated by applying the temperature scalar to spatial coordinates. Search for and mark heat dissipation dead zones in the temperature and flow field distribution inside the cabin; Calculate the spatial coordinates of the geometric center of the heat dissipation dead zone, construct derivative-enhanced constraints, calculate the sum of the horizontal and vertical second-order partial derivatives of the temperature scalar with respect to the spatial coordinates, divide the sum by the first-order derivative of the temperature scalar with respect to time to obtain the thermal diffusivity, and use the absolute value of the difference between the thermal diffusivity and the preset standard thermal diffusivity as the residual value. Determine if the residual value is less than the preset residual threshold. If it is, output the spatial coordinates of the heat dissipation dead zone. Otherwise, use the residual value to calculate the gradient and update the network parameters in reverse until the residual value meets the condition.

4. The offshore wind turbine nacelle transformer heat dissipation collaborative optimization system according to claim 1, characterized in that, The working process of the temperature rise prediction and classification module includes: The load current data at the current moment is extracted from the multidimensional state parameter set, and the product of the square value of the load current data and the DC resistance value of the transformer winding is calculated to obtain the heat generation power value at the current moment. The corresponding airflow velocity scalar is extracted based on the spatial coordinates of the heat dissipation dead zone. The difference between the average airflow velocity scalar inside the computer cabin and the airflow velocity scalar inside the heat dissipation dead zone is divided by the average airflow velocity scalar inside the cabin to obtain the heat dissipation drag coefficient. The product of the current heat generation power and the heat dissipation resistance coefficient is calculated as the instantaneous heat accumulation factor. The winding hot spot temperature data is extracted from the multidimensional state parameter set, and the slope value that changes with time is calculated as the historical temperature rise trend factor. The summation with the instantaneous heat accumulation factor is used to obtain the predicted value of the temperature rise rate. The predicted temperature rise rate is compared with the preset temperature rise rate threshold, and the heat dissipation requirement level is output.

5. The offshore wind turbine nacelle transformer heat dissipation collaborative optimization system according to claim 1, characterized in that, The working process of the collaborative control decision module includes: The Actor network of the improved TD3 algorithm is constructed by concatenating the predicted temperature rise rate of change with the heat dissipation demand level into a one-dimensional state vector input. The one-dimensional state vector is linearly weighted and summed to obtain the hidden layer input value. The hidden layer input value is then transformed by a nonlinear activation function to obtain the hidden layer feature value. The above linear weighting and nonlinear activation operations are iteratively executed through three fully connected layers to output the air duct opening adjustment amount, cooling fan speed command and phase change material activation signal, and combine them to generate tentative control actions. The tentative control actions are input into the safety critique network, which calculates and outputs a safety correction factor. A one-dimensional state vector is concatenated with an exploratory control action to form a combined input vector, which is then input into the first and second Critic subnetworks of the dual Critic network. In the first and second Critic subnetworks, the combined input vector is subjected to linear weighted summation and nonlinear activation operation to obtain the first state action value score and the second state action value score. The numerical values ​​of the first state action value score and the second state action value score are compared, and the smaller value is selected as the target value evaluation result. The composite loss function value is obtained by summing the negative value of the target value assessment result and the safety correction factor. The first derivative of the composite loss function value with respect to the Actor network parameters is calculated as the gradient value. The updated Actor network parameter values ​​are obtained by subtracting the product of the gradient value and the preset learning rate from the Actor network parameter values. The network parameter update step is repeated until the composite loss function value is less than the preset convergence threshold. The tentative control action output by the Actor network is used as the collaborative control instruction set.

6. The offshore wind turbine nacelle transformer heat dissipation collaborative optimization system according to claim 5, characterized in that, The process of inputting tentative control actions into the security critique network, calculating and outputting a security correction factor includes: The tentative control actions are input into the safety critique network, and the values ​​of the air duct opening adjustment, cooling fan speed command and phase change material activation signal in the tentative control actions are extracted respectively. Each value is judged to see if it exceeds the preset safety action threshold range. When the value is greater than the upper limit of the preset safety action threshold range, the absolute value of the difference between the value and the upper limit is calculated and multiplied by the preset penalty weight coefficient. When the value is less than the lower limit of the preset safety action threshold range, the absolute value of the difference between the lower limit and the value is calculated and multiplied by the preset penalty weight coefficient. When the value is within the preset safety action threshold range, the penalty item is set to zero. The safety correction factor is obtained by summing the three penalty items corresponding to the air duct opening adjustment amount, the cooling fan speed command, and the phase change material activation signal.

7. The offshore wind turbine nacelle transformer heat dissipation collaborative optimization system according to claim 1, characterized in that, The operation process of the instruction execution module includes: Read the collaborative control instruction set, calculate the target baffle angle value based on the air duct opening adjustment amount, send the rotation angle command, drive the baffle to rotate to the target angle value, and guide the cooling airflow to the spatial coordinate position of the heat dissipation dead zone. Calculate the corresponding target power supply frequency value based on the cooling fan speed command, adjust the output pulse frequency of the fan driver to the target power supply frequency value, and change the fan motor speed to the speed value required by the cooling fan speed command. When the phase change material activation signal is high, the solenoid valve switch of the phase change material storage device is closed, releasing the phase change material so that it flows through the heat dissipation dead zone to absorb heat; when the phase change material activation signal is low, the solenoid valve switch is opened. The system collects the temperature and flow field distribution inside the cabin after the data collection process, identifies the temperature changes at the spatial coordinates of the heat dissipation dead zone, and determines that the directional cooling effect for the heat dissipation dead zone is achieved when the rate of temperature decrease is greater than the preset cooling threshold. At the same time, the system continuously collects the temperature data of the winding hot spots at preset time intervals, calculates the difference between the temperature data of the winding hot spots at two adjacent collection times to obtain the temperature change difference, and determines that the peak heat load transfer result is achieved when the temperature change difference is less than or equal to zero.

8. The offshore wind turbine nacelle transformer heat dissipation collaborative optimization system according to claim 1, characterized in that, The operation process of the security protection module includes: Collect real-time winding temperature data after the execution of the coordinated control instruction set, compare the real-time winding temperature data with the preset maximum temperature threshold, and determine that it exceeds the safe range when it is greater than the preset maximum temperature threshold. When the load is determined to be outside the safe range, a voltage reduction command is sent to the on-load tap changer of the transformer to gradually reduce the transformer tap position. The load current data is monitored in real time, and the ratio of the current load current data to the rated current value is calculated and multiplied by the rated capacity of the transformer to obtain the limited transformer load capacity value.

Images