Method and system for optimizing heat dissipation power of hydraulic cooler based on air flow

Abstract

The invention provides a hydraulic cooler heat dissipation power optimization method and system based on air flow, and relates to the technical field of hydraulic system heat dissipation, and the method comprises the steps: obtaining real-time operation parameters, and calculating the actual heat dissipation power; obtaining a target fan rotating speed of the maximum predicted heat dissipation power based on an improved particle swarm algorithm, and when a prediction error is greater than a threshold value, optimizing the model by adopting a deep reinforcement learning algorithm; and a fan rotating speed closed-loop control system is established based on sliding mode variable structure control, and the switching function gain coefficient is adaptively adjusted according to the environment temperature change rate, so that the fan rotating speed is dynamically adjusted. The cooling efficiency of the hydraulic cooler is improved, and the service life of equipment is prolonged.

Description

Technical Field

[0001] This invention relates to the field of hydraulic system heat dissipation technology, and in particular to a method and system for optimizing the heat dissipation power of a hydraulic cooler based on airflow. Background Technology

[0002] Hydraulic systems are widely used in industrial equipment, construction machinery, and aerospace, and hydraulic coolers, as key components of hydraulic systems, play a crucial role in maintaining system temperature stability and ensuring normal operation. When a hydraulic system is working, the hydraulic oil flowing under high pressure generates a large amount of heat. If heat is not dissipated in time, the excessively high hydraulic oil temperature will lead to a decrease in oil viscosity, aging of seals, and accelerated wear of components, seriously affecting system performance and service life.

[0003] Traditional hydraulic cooler heat dissipation control typically employs simple on / off control or fixed-speed control, with the fan either running at full speed or operating at a preset fixed speed. While this method is simple in structure and easy to implement, it is difficult to flexibly adjust the heat dissipation capacity according to actual working conditions and environmental conditions, easily leading to energy waste or insufficient heat dissipation. As hydraulic systems develop towards higher efficiency, precision, and intelligence, higher demands are placed on the heat dissipation efficiency and control accuracy of cooling systems. In recent years, some researchers have begun to apply methods such as fuzzy control and PID control to hydraulic cooling systems to achieve more precise temperature control. However, when facing complex working conditions and changing environments, the adaptive ability and optimization effect of these methods remain limited. Summary of the Invention

[0004] The present invention provides a method and system for optimizing the heat dissipation power of a hydraulic cooler based on airflow, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides a method for optimizing the heat dissipation power of a hydraulic cooler based on airflow, comprising: The real-time operating parameters of the hydraulic cooler are obtained, the current heat dissipation power of the hydraulic cooler is calculated based on the real-time operating parameters, and the current heat dissipation power is used as the actual heat dissipation power of the hydraulic cooler. The real-time operating parameters are input into the hydraulic cooler heat dissipation power prediction model. The target fan speed that maximizes the predicted heat dissipation power is calculated based on the improved particle swarm optimization algorithm. The improved particle swarm optimization algorithm introduces an adaptive adjustment strategy for inertial weights and a dynamic convergence factor on the basis of the original particle swarm optimization algorithm. When the power error between the predicted heat dissipation power of the hydraulic cooler heat dissipation power prediction model and the actual heat dissipation power is greater than a preset power threshold, an optimization objective function is constructed based on a deep reinforcement learning algorithm. The power error is used as the reward function, and the network parameters of the hydraulic cooler heat dissipation power prediction model are iteratively optimized online using the policy gradient method until the power error is less than the preset power threshold. A closed-loop control system for wind turbine speed is established based on a sliding mode variable structure control algorithm. A sliding mode switching surface is constructed according to the deviation between the target wind turbine speed and the current wind turbine speed. An adaptive reaching law is used to achieve rapid tracking control of the wind turbine speed. Specifically, the exponential coefficient of the adaptive reaching law is adjusted according to the dynamic change of the deviation. At the same time, the ambient temperature change rate is collected and calculated in real time. The gain coefficient of the switching function of the sliding mode variable structure control algorithm is adaptively adjusted according to the ambient temperature change rate to form a composite adaptive control law, thereby realizing the dynamic adjustment of the wind turbine speed.

[0006] An optimization objective function is constructed based on a deep reinforcement learning algorithm, and the power error is used as the reward function. The network parameters of the hydraulic cooler heat dissipation power prediction model are then iteratively optimized online using a policy gradient method, including: A multi-objective reward function is constructed, which includes three weighted product terms. The first weighted product term is the product of the first weight coefficient and the exponential decay value of the power error. The second weighted product term is the product of the second weight coefficient and the exponential decay value of the power error change rate. The third weighted product term is the product of the third weight coefficient and the prediction smoothness evaluation value. The policy gradient optimization objective is constructed based on the multi-objective reward function. The expected value of the product of the log-likelihood value of the policy network parameters and the discount reward is used as the policy gradient calculation formula. The policy gradient value is calculated by the policy gradient calculation formula, wherein the discount reward is calculated by the multi-objective reward function. The dynamic learning rate is calculated based on the value estimate of the current state, where the value estimate represents the expected cumulative reward in the current state, and the dynamic learning rate is the product of the initial learning rate and the exponential decay function of the value estimate. The network parameter adjustment amount is obtained by multiplying the policy gradient value by the dynamic learning rate. The network parameter adjustment amount is then used to update the current network parameters to obtain the updated network parameters. The magnitude of the network parameter adjustment amount is adaptively adjusted according to the value estimate. The updated network parameters are subjected to anomaly detection, and the final network parameters are determined using a parameter rollback mechanism and updated in the hydraulic cooler heat dissipation power prediction model.

[0007] The target fan speed that maximizes the predicted heat dissipation power is calculated based on an improved particle swarm optimization (PSO) algorithm. This improved PSO algorithm introduces an adaptive inertia weight adjustment strategy and a dynamic convergence factor, based on the original PSO algorithm, including: An improved particle swarm optimization algorithm is constructed by adaptively adjusting the inertia weight. The difference between the maximum and minimum inertia weights is multiplied by an exponential decay function based on the current iteration number. The product is then subtracted from the maximum inertia weight to obtain the inertia weight for the current iteration. The improved particle swarm optimization algorithm is constructed using a dynamic convergence factor, which includes a cognitive learning factor for adjusting the influence of the individual particle's historical best position on the particle's motion, and a social learning factor for adjusting the influence of the swarm's historical best position on the particle's motion. The product of the inertial weight and the particle's current velocity, the product of the cognitive learning factor and the individual optimal position deviation, the product of the social learning factor and the global optimal position deviation, and the weighted sum of the flight disturbance term are used as the particle's velocity update. Calculate a population diversity index, construct an adaptive mutation probability using the population diversity index, wherein the adaptive mutation probability decreases exponentially as the population diversity index increases, and perform mutation operations on the particle swarm based on the adaptive mutation probability. The mutated particle positions are substituted into the heat dissipation power prediction model to obtain the corresponding predicted heat dissipation power. Based on the predicted heat dissipation power, fitness evaluation is performed. The particle position with the best evaluation result is saved to the elite solution set, and the elite solution set is dynamically updated in each iteration. The wind turbine speed corresponding to the position of the particle with the best fitness in the elite solution set is taken as the target wind turbine speed.

[0008] The method further includes: The cognitive learning factor is calculated by adding the product of the first random initial value and the square of the ratio of the current iteration number to the maximum iteration number; The social learning factor is calculated by adding the product of the second random initial value and the square of the ratio of the current iteration number to the maximum iteration number; The individual optimal position deviation is the difference between the individual particle's historical optimal position and its current position, while the global optimal position deviation is the difference between the collective particle population's historical optimal position and its current position.

[0009] The exponential coefficient of the adaptive reaching law is adjusted according to the dynamic change of the deviation. Simultaneously, the rate of change of ambient temperature is collected and calculated in real time. The gain coefficient of the switching function of the sliding mode variable structure control algorithm is adaptively adjusted according to the rate of change of ambient temperature, forming a composite adaptive control law to achieve dynamic adjustment of the fan speed, including: The deviation change rate is calculated based on the deviation, and the sliding surface of the sliding mode variable structure control algorithm is constructed based on the deviation and the deviation change rate. The sliding surface is the product of the deviation and a preset constant plus the deviation change rate. The absolute value of the deviation and the absolute value of the deviation change rate are weighted and summed to obtain a comprehensive dynamic characteristic quantity. Based on the comprehensive dynamic characteristic quantity, the exponential coefficient of the adaptive reaching law is constructed. The exponential coefficient of the adaptive reaching law is formed by adding the basic exponential value to a negative exponential function with the comprehensive dynamic characteristic quantity as the independent variable. The ambient temperature is differentially calculated to obtain the rate of temperature change, and the switching function gain coefficient is designed based on the absolute value of the rate of temperature change. The switching function gain coefficient is a linear combination of the base gain value and the absolute value of the rate of temperature change. A first control term is constructed based on the exponential coefficients of the adaptive reaching law. The first control term includes the product of a variable exponential term of the absolute value of the sliding surface and a sign function, wherein the exponent of the variable exponential term is determined by the exponential coefficients of the adaptive reaching law. A second control term is constructed based on the gain coefficient of the switching function, and the second control term is the product of the gain coefficient of the switching function and the sign function of the sliding surface; The first control term and the second control term are superimposed to obtain a composite adaptive control law, and the composite adaptive control law is output as a control command to the fan execution system to realize the dynamic adjustment of the fan speed.

[0010] After achieving dynamic adjustment of the fan speed, the method further includes: The inlet temperature in the real-time operating parameters is compared with the preset temperature threshold. When the inlet temperature is detected to be greater than the preset temperature threshold, the current real-time operating parameters are input into the hydraulic cooler heat dissipation power prediction model to calculate the target flow rate that can reduce the inlet temperature to below the preset temperature threshold. The pressure compensation correction value of the hydraulic system is calculated based on the target flow rate. The pressure compensation correction value is then superimposed on the current pressure compensation value of the hydraulic system. Through closed-loop control of pressure compensation, the hydraulic flow rate is gradually adjusted to the target flow rate, thereby achieving optimized control of the heat dissipation power of the hydraulic cooler.

[0011] A second aspect of the present invention provides a hydraulic cooler heat dissipation power optimization system for airflow, comprising: The first unit is used to acquire the real-time operating parameters of the hydraulic cooler, calculate the current heat dissipation power of the hydraulic cooler based on the real-time operating parameters, and use the current heat dissipation power as the actual heat dissipation power of the hydraulic cooler. The second unit is used to input the real-time operating parameters into the hydraulic cooler heat dissipation power prediction model, and calculate the target fan speed that maximizes the predicted heat dissipation power based on the improved particle swarm optimization algorithm. The improved particle swarm optimization algorithm introduces an adaptive adjustment strategy for inertial weights and a dynamic convergence factor on the basis of the original particle swarm optimization algorithm. When the power error between the predicted heat dissipation power of the hydraulic cooler heat dissipation power prediction model and the actual heat dissipation power is greater than a preset power threshold, an optimization objective function is constructed based on a deep reinforcement learning algorithm. The power error is used as the reward function, and the network parameters of the hydraulic cooler heat dissipation power prediction model are iteratively optimized online using the policy gradient method until the power error is less than the preset power threshold. The third unit is used to establish a closed-loop control system for wind turbine speed based on the sliding mode variable structure control algorithm. It constructs a sliding mode switching surface based on the deviation between the target wind turbine speed and the current wind turbine speed, and uses an adaptive approach law to achieve rapid tracking control of the wind turbine speed. Specifically, the exponential coefficient of the adaptive approach law is adjusted according to the dynamic changes in the deviation, while the ambient temperature change rate is collected and calculated in real time. The gain coefficient of the switching function of the sliding mode variable structure control algorithm is adaptively adjusted according to the ambient temperature change rate to form a composite adaptive control law, thereby achieving dynamic adjustment of the wind turbine speed.

[0012] A third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0013] Fourth aspect of the present invention, A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0014] This invention provides a method for optimizing the heat dissipation power of a hydraulic cooler based on airflow, which has the following beneficial effects: By introducing an improved particle swarm optimization algorithm to calculate the target fan speed, this algorithm adopts an adaptive adjustment strategy of inertial weight and a dynamic convergence factor, which improves search efficiency, avoids local optimum traps, and can more accurately find the fan operating point that maximizes heat dissipation power, thereby improving the system's energy efficiency ratio.

[0015] An optimization objective function is constructed based on a deep reinforcement learning algorithm, and the power error is used as the reward function. The policy gradient method is used to perform online iterative optimization of the heat dissipation power prediction model, which enables the model to have adaptive learning ability and continuously optimize the model accuracy according to the actual operating state, thereby improving the system's adaptability under complex operating conditions.

[0016] A closed-loop control system for the fan speed is established by using a sliding mode variable structure control algorithm, and an adaptive adjustment mechanism based on the rate of change of ambient temperature is introduced to form a composite adaptive control law. This control method has strong anti-interference ability, can quickly respond to temperature fluctuations, achieve precise adjustment of fan speed, ensure stable operation of the hydraulic system under variable working conditions, and effectively extend the service life of the equipment. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the method for optimizing the heat dissipation power of a hydraulic cooler based on airflow according to an embodiment of the present invention. Figure 2 A schematic diagram of the improved particle swarm optimization algorithm architecture. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The technical solution of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes will not be repeated in some embodiments.

[0020] Figure 1 This is a flowchart illustrating the method for optimizing the heat dissipation power of a hydraulic cooler based on airflow according to an embodiment of the present invention. Figure 1 As shown, the method includes: The real-time operating parameters of the hydraulic cooler are obtained, the current heat dissipation power of the hydraulic cooler is calculated based on the real-time operating parameters, and the current heat dissipation power is used as the actual heat dissipation power of the hydraulic cooler. The real-time operating parameters are input into the hydraulic cooler heat dissipation power prediction model. The target fan speed that maximizes the predicted heat dissipation power is calculated based on the improved particle swarm optimization algorithm. The improved particle swarm optimization algorithm introduces an adaptive adjustment strategy for inertial weights and a dynamic convergence factor on the basis of the original particle swarm optimization algorithm. When the power error between the predicted heat dissipation power of the hydraulic cooler heat dissipation power prediction model and the actual heat dissipation power is greater than a preset power threshold, an optimization objective function is constructed based on a deep reinforcement learning algorithm. The power error is used as the reward function, and the network parameters of the hydraulic cooler heat dissipation power prediction model are iteratively optimized online using the policy gradient method until the power error is less than the preset power threshold. A closed-loop control system for wind turbine speed is established based on a sliding mode variable structure control algorithm. A sliding mode switching surface is constructed according to the deviation between the target wind turbine speed and the current wind turbine speed. An adaptive reaching law is used to achieve rapid tracking control of the wind turbine speed. Specifically, the exponential coefficient of the adaptive reaching law is adjusted according to the dynamic change of the deviation. At the same time, the ambient temperature change rate is collected and calculated in real time. The gain coefficient of the switching function of the sliding mode variable structure control algorithm is adaptively adjusted according to the ambient temperature change rate to form a composite adaptive control law, thereby realizing the dynamic adjustment of the wind turbine speed.

[0021] In one optional implementation, an optimization objective function is constructed based on a deep reinforcement learning algorithm, with the power error used as the reward function. The network parameters of the hydraulic cooler heat dissipation power prediction model are then iteratively optimized online using a policy gradient method, including: A multi-objective reward function is constructed, which includes three weighted product terms. The first weighted product term is the product of the first weight coefficient and the exponential decay value of the power error. The second weighted product term is the product of the second weight coefficient and the exponential decay value of the power error change rate. The third weighted product term is the product of the third weight coefficient and the prediction smoothness evaluation value. The policy gradient optimization objective is constructed based on the multi-objective reward function. The expected value of the product of the log-likelihood value of the policy network parameters and the discount reward is used as the policy gradient calculation formula. The policy gradient value is calculated by the policy gradient calculation formula, wherein the discount reward is calculated by the multi-objective reward function. The dynamic learning rate is calculated based on the value estimate of the current state, where the value estimate represents the expected cumulative reward in the current state, and the dynamic learning rate is the product of the initial learning rate and the exponential decay function of the value estimate. The network parameter adjustment amount is obtained by multiplying the policy gradient value by the dynamic learning rate. The network parameter adjustment amount is then used to update the current network parameters to obtain the updated network parameters. The magnitude of the network parameter adjustment amount is adaptively adjusted according to the value estimate. The updated network parameters are subjected to anomaly detection, and the final network parameters are determined using a parameter rollback mechanism and updated in the hydraulic cooler heat dissipation power prediction model.

[0022] This invention provides a method for constructing an optimization objective function based on a deep reinforcement learning algorithm. This method uses power error as the reward function and employs a policy gradient method to perform online iterative optimization of the network parameters of a hydraulic cooler heat dissipation power prediction model.

[0023] In the specific implementation process, a multi-objective reward function is first constructed, which consists of three weighted product terms. The first weighted product term is the product of the first weight coefficient and the exponentially decaying value of the power error, expressed as α multiplied by exp(-β|Pactual - PPredicted|), where α is set to 0.6, β is set to 0.05, Pactual represents the actual heat dissipation power, and PPredicted represents the predicted heat dissipation power. The second weighted product term is the product of the second weight coefficient and the exponentially decaying value of the power error change rate, expressed as γ multiplied by exp(-δ|(Pactual,tPPredicted,t)-(Pactual,t-1-PPredicted,t-1)| / |Pactual,t-1-PPredicted,t-1|), where γ is set to 0.3, δ is set to 0.1, t represents the current time, and t-1 represents the previous time. The third weighted product term is the product of the third weight coefficient and the prediction smoothness evaluation value, expressed as η multiplied by (1-|P prediction, tP prediction, t-1| / P prediction, t-1), where η is set to 0.1. The sum of these three weighted product terms constitutes the complete multi-objective reward function, i.e., R = α·exp(-β|P actual - P prediction|) + γ·exp(-δ|(P actual, tP prediction, t)-(P actual, t-1 - P prediction, t-1)| / |P actual, t-1 - P prediction, t-1|) + η·(1-|P prediction, tP prediction, t-1| / P prediction, t-1).

[0024] Based on the aforementioned multi-objective reward function, a policy gradient optimization objective is constructed. The expected value of the product of the log-likelihood of the policy network parameters and the discounted reward is used as the formula for calculating the policy gradient. In practice, for each time step t, the log probability ln(π(at|st;θ)) of the action at (i.e., prediction adjustment) in state st is calculated, where θ represents the current policy network parameters. The discounted reward Gt is calculated as the sum of discounts on all future rewards starting from the current time t, i.e., Gt = Rt + λRt+1 + λ²Rt+2 + ..., where λ is the discount factor, set to 0.95. The policy gradient value is calculated as the product of the log-likelihood and the discounted reward, i.e., ∇θJ(θ) = ln(π(at|st;θ)) · Gt. In practice, multiple sampling is used to obtain the action and state sequences, and then the average gradient value is calculated to reduce variance. For example, 10 samplings are performed, each containing 100 time steps, and then the average gradient value is calculated.

[0025] Next, the dynamic learning rate is calculated based on the value estimate of the current state. The value estimate V(st) represents the expected cumulative reward in the current state st and is calculated using the value network. The dynamic learning rate is defined as the product of the initial learning rate and the exponential decay function of the value estimate, i.e., η(st) = η0 · exp(-ρ·V(st)), where η0 is the initial learning rate, set to 0.01, and ρ is the decay coefficient, set to 0.1. The value estimate V(st) is calculated using the value network, which takes the current system state st as input and outputs a scalar value representing the expected cumulative reward. The value network uses a three-layer fully connected neural network with 64 and 32 hidden layer nodes, respectively, and ReLU activation function.

[0026] After calculating the dynamic learning rate, the policy gradient value is multiplied by the dynamic learning rate to obtain the network parameter adjustment amount, i.e., Δθ = η(st) · ∇θJ(θ). This network parameter adjustment amount is then used to update the current network parameters, i.e., θnew = θold + Δθ, where θnew represents the updated network parameters and θold represents the original network parameters. It is worth noting that the magnitude of the network parameter adjustment amount adaptively adjusts with the value estimate. When the system is in a high-value state, the value estimate V(st) is large, resulting in a smaller dynamic learning rate and a smaller parameter update magnitude, maintaining system stability. When the system is in a low-value state, the value estimate V(st) is small, the dynamic learning rate is large, the parameter update magnitude is large, and the system optimization speed is accelerated.

[0027] Anomaly detection is performed on the updated network parameters, and a parameter rollback mechanism is used to determine the final network parameters. Anomaly detection includes two aspects: parameter magnitude detection and prediction effect detection. Parameter magnitude detection determines whether the updated parameters exceed a reasonable range, i.e., checking whether |θnew,i - θold,i| / |θold,i| > 0.2 holds true, where i represents the parameter index. Prediction effect detection determines whether the prediction error using the updated parameters has significantly increased, i.e., checking whether |Pactual - PPredicted,new| / |Pactual - PPredicted,old| > 1.5 holds true. If either detection result is anomaly, the parameter rollback mechanism is triggered, halving the parameter update magnitude, i.e., θnew = θold + 0.5·Δθ, and then anomaly detection is performed again, repeating up to 3 times. If anomalies are detected in all 3 iterations, the update is abandoned, and the original parameters remain unchanged. After determining the final network parameters, they are updated into the hydraulic cooler heat dissipation power prediction model, completing one parameter optimization iteration.

[0028] In practical applications, this method performs parameter optimization every 5 minutes, using historical data from the most recent 30 minutes as training samples. After applying this method to a real cooling system, the average error of heat dissipation power prediction decreased from 8.7% to 3.2%, the prediction stability was significantly improved, and the energy efficiency was increased by 12.5%.

[0029] In one optional implementation, the target fan speed that maximizes the predicted heat dissipation power is calculated based on an improved particle swarm optimization (PSO) algorithm. The improved PSO algorithm introduces an adaptive inertia weight adjustment strategy and a dynamic convergence factor on top of the original PSO algorithm, including: An improved particle swarm optimization algorithm is constructed by adaptively adjusting the inertia weight. The difference between the maximum and minimum inertia weights is multiplied by an exponential decay function based on the current iteration number. The product is then subtracted from the maximum inertia weight to obtain the inertia weight for the current iteration. The improved particle swarm optimization algorithm is constructed using a dynamic convergence factor, which includes a cognitive learning factor for adjusting the influence of the individual particle's historical best position on the particle's motion, and a social learning factor for adjusting the influence of the swarm's historical best position on the particle's motion. The product of the inertial weight and the particle's current velocity, the product of the cognitive learning factor and the individual optimal position deviation, the product of the social learning factor and the global optimal position deviation, and the weighted sum of the flight disturbance term are used as the particle's velocity update. Calculate a population diversity index, construct an adaptive mutation probability using the population diversity index, wherein the adaptive mutation probability decreases exponentially as the population diversity index increases, and perform mutation operations on the particle swarm based on the adaptive mutation probability. The mutated particle positions are substituted into the heat dissipation power prediction model to obtain the corresponding predicted heat dissipation power. Based on the predicted heat dissipation power, fitness evaluation is performed. The particle position with the best evaluation result is saved to the elite solution set, and the elite solution set is dynamically updated in each iteration. The wind turbine speed corresponding to the position of the particle with the best fitness in the elite solution set is taken as the target wind turbine speed.

[0030] In this embodiment, a method is proposed to calculate the target fan speed that maximizes the predicted heat dissipation power based on an improved particle swarm optimization (PSO) algorithm. This method optimizes the performance of the traditional PSO algorithm by introducing an adaptive adjustment strategy for inertia weights and a dynamic convergence factor.

[0031] The improved particle swarm optimization algorithm first constructs an adaptive adjustment strategy for inertia weight. In practice, the maximum inertia weight is set to 0.9, the minimum inertia weight to 0.4, the current iteration number to t, and the maximum iteration number to 100. The exponential decay function can be expressed as the current iteration number t divided by the square of the maximum iteration number. This design causes the inertia weight to decrease slowly in the early stages of iteration and rapidly in the later stages. For example, when the iteration reaches 20, the exponential decay function value is 0.04, and the inertia weight is calculated as 0.9 - (0.9 - 0.4) × 0.04 = 0.88, which is beneficial for the algorithm's global search. When the iteration reaches 80, the exponential decay function value is 0.64, and the inertia weight is calculated as 0.9 - (0.9 - 0.4) × 0.64 = 0.58, at which point the algorithm focuses more on local search.

[0032] Regarding the construction of the dynamic convergence factors, the cognitive learning factor and the social learning factor control the degree to which particles follow their individual and group historical best positions, respectively. The cognitive learning factor is set to a relatively large value of 2.5 in the early stages of iteration, decreasing linearly to 1.5 as iterations progress; the social learning factor, on the other hand, increases linearly from an initial value of 1.5 to 2.5. This design allows particles to rely more on their own experience to explore the solution space in the early stages of the search, while later they utilize more group information to converge to the optimal solution. Specifically, when the 40th iteration is reached, the cognitive learning factor is 2.1 and the social learning factor is 1.9, indicating that particles are gradually shifting from individual exploration to group collaboration.

[0033] The calculation of particle velocity update considers four parts: inertia term, cognitive term, social term, and perturbation term. The inertia term is the product of the current inertia weight and the current particle velocity; the cognitive term is the product of the cognitive learning factor, a random number (between 0 and 1), and the individual's optimal position deviation; the social term is the product of the social learning factor, a random number (between 0 and 1), and the global optimal position deviation; the perturbation term is a small-amplitude random perturbation with a value range of [-0.01, 0.01]. For example, in the search for wind turbine rotation speed, when the particle's current velocity is 2 rpm, current position is 800 rpm, individual historical best position is 850 rpm, group historical best position is 870 rpm, inertia weight is 0.7, cognitive learning factor is 2.0, social learning factor is 2.0, random numbers are 0.6 and 0.8 respectively, and perturbation is 0.005, the velocity update is calculated as 0.7×2+2.0×0.6×(850-800)+2.0×0.8×(870-800)+0.005×800=1.4+60+112+4=177.4 rpm.

[0034] Population diversity index is calculated based on the variance of particle positions. In practice, for a population of 30 particles, the sum of squared Euclidean distances from all particle positions to the average position is calculated, and then divided by the number of particles to obtain the population diversity index. For example, when all particles are relatively concentrated, the calculated population diversity index can be 10; when particles are dispersed, the index can reach 100 or higher.

[0035] The adaptive mutation probability decreases exponentially with increasing population diversity index. Its basic form is the base mutation probability multiplied by the negative exponent of the natural base e, where the exponent is the population diversity index divided by the control parameter. Setting the base mutation probability to 0.1 and the control parameter to 50, when the population diversity index is 10, the mutation probability is 0.1 × e. (-10 / 50) =0.082; when the diversity index is 100, the probability of variation decreases to 0.1×e (-100 / 50) =0.014. The mutation operation judges each particle based on this probability. If mutation is required, a random perturbation is added or removed from the current position. The perturbation size is 1%-5% of the search space.

[0036] The particle positions are substituted into the heat dissipation power prediction model for evaluation. The heat dissipation power prediction model is trained based on historical data; the input is the fan speed, and the output is the predicted heat dissipation power. For example, when the fan speed is 800 rpm, the predicted heat dissipation power is 100W; at 900 rpm, the predicted heat dissipation power is 120W; and at 1000 rpm, the predicted heat dissipation power is 115W. Fitness evaluation directly uses the predicted heat dissipation power value; the higher the value, the higher the fitness.

[0037] The elite solution set stores the best solutions found in each iteration, with a capacity of 5. In the initial iteration, the first 5 best solutions are added to the set. In subsequent iterations, if a new solution better than the worst solution in the set is found, the worst solution is replaced. For example, if the fan speeds and corresponding heat dissipation powers stored in the initial iteration elite solution set are {(920rpm, 123W), (910rpm, 122W), (930rpm, 121W), (900rpm, 120W), (890rpm, 119W)}, and in a new iteration it is found that a speed of 950rpm corresponds to a heat dissipation power of 124W, then (890rpm, 119W) in the set is replaced.

[0038] After 100 iterations, the solution with the highest fitness is selected from the elite solution set as the final result. If the final elite solution set is {(950rpm,124W), (920rpm,123W), (910rpm,122W), (930rpm,121W),(900rpm,120W)}, then the target fan speed is determined to be 950rpm and the predicted heat dissipation power is 124W.

[0039] Figure 2 This diagram illustrates the improved particle swarm optimization (PSO) algorithm architecture, covering the entire process from initialization to result output. The invention divides the entire algorithm flow into four key stages: initialization, adaptive adjustment, iterative optimization, and result output. In the initialization stage, particle swarm initialization, parameter initialization, prediction model loading, and elite solution initialization are completed. The adaptive adjustment stage dynamically adjusts inertia weights, cognitive learning factors, and social learning factors, as well as calculates population diversity and constructs adaptive mutation probabilities. The iterative optimization stage includes core functions such as particle velocity update, position update, mutation operation, boundary handling, heat dissipation power prediction, fitness evaluation, individual / global optimal update, and elite solution update. Finally, in the result output stage, the target wind turbine rotational speed corresponding to the optimal particle position is output. This invention significantly improves the convergence speed and solution quality of the algorithm through multi-parameter adaptive adjustment and elite solution preservation strategies, avoiding the pitfalls of traditional PSO algorithms that easily get trapped in local optima, making it particularly suitable for the field of wind turbine heat dissipation power optimization control.

[0040] In one optional implementation, the method further includes: The cognitive learning factor is calculated by adding the product of the first random initial value and the square of the ratio of the current iteration number to the maximum iteration number; The social learning factor is calculated by adding the product of the second random initial value and the square of the ratio of the current iteration number to the maximum iteration number; The individual optimal position deviation is the difference between the individual particle's historical optimal position and its current position, while the global optimal position deviation is the difference between the collective particle population's historical optimal position and its current position.

[0041] This embodiment describes in detail a method for calculating cognitive learning factors and social learning factors. This method not only provides the calculation process for cognitive and social learning factors, but also defines the specific calculation methods for individual optimal position deviation and global optimal position deviation.

[0042] The cognitive learning factor of this method is obtained through a series of calculation steps. Specifically, the system first sets a first random initial value for the cognitive learning factor. This initial value can be a value randomly generated within a preset range, such as randomly selecting a value between 1.5 and 2.5 as the initial value of the cognitive learning factor, for example, 2.0. Simultaneously, the system also sets a change amount for the cognitive learning factor, which represents the maximum change in the cognitive learning factor throughout the entire iteration process; for example, it can be set to -0.5. During the algorithm's iteration process, the system calculates the ratio of the current iteration number to the maximum iteration number. For example, if the current iteration is the 50th iteration and the maximum iteration number is 200, then this ratio is 50 / 200 = 0.25. The system then calculates the square of this ratio, which is 0.25 in this example. 2 =0.0625. Then, the system multiplies the change in the cognitive learning factor by this squared value to obtain the current change, which in this example is -0.5 × 0.0625 = -0.03125. Finally, the system adds this change to the first random initial value to obtain the cognitive learning factor for the current iteration, which in this example is 2.0 + (-0.03125) = 1.96875. This calculation method allows the cognitive learning factor to gradually change with the number of iterations, enabling the algorithm to gradually transition from emphasizing individual exploration to emphasizing group cooperation during the search process.

[0043] The calculation process for the social learning factor is similar to that of the cognitive learning factor, but it uses different initial values ​​and variables. The system sets a second random initial value for the social learning factor; for example, a value randomly selected between 1.5 and 2.5, such as 2.0, can be used. Simultaneously, the system sets a variable for the social learning factor, such as 0.5. During the algorithm's iteration, the system also calculates the ratio of the current iteration number to the maximum iteration number and its square. For example, in the 50th iteration, the square of this ratio is 0.0625. Then, the system multiplies the variable for the social learning factor by this squared value to obtain the current variable, which in this example is 0.5 × 0.0625 = 0.03125. Finally, the system adds this variable to the second random initial value to obtain the social learning factor for the current iteration, which in this example is 2.0 + 0.03125 = 2.03125. This calculation method allows the social learning factor to gradually increase with the number of iterations, enabling the algorithm to gradually increase its focus on the global optimal position during the search process.

[0044] In practice, the individual optimal position deviation refers to the difference between a particle's historical best position and its current position. For example, suppose a particle's current position in two-dimensional space is (3, 4), while its historical best position is (5, 7). Then the individual optimal position deviation is (5, 7) - (3, 4) = (2, 3). This deviation vector represents the direction and distance the particle needs to move towards its historical best position, and it is an important parameter when calculating particle velocity updates.

[0045] The global optimal position deviation refers to the difference between the historical optimal position of the particle swarm and its current position. For example, suppose the global optimal position of the entire particle swarm is (8, 9), and the current position of a certain particle is (3, 4). Then, for this particle, the global optimal position deviation is (8, 9) - (3, 4) = (5, 5). This deviation vector represents the direction and distance the particle needs to move towards the historical optimal position of the swarm, and it is also an important parameter when calculating particle velocity updates.

[0046] In practical applications, the above method can be applied to various optimization problems. For example, in a function optimization problem, the system can initialize a swarm of particles, each representing a candidate solution in the solution space. By calculating the cognitive learning factor and social learning factor using the above method, and combining the individual optimal position deviation and the global optimal position deviation, the system can effectively update the particle's velocity and position, guiding the particles to move towards a better solution space region. As the number of iterations increases, the cognitive learning factor gradually decreases, while the social learning factor gradually increases, causing the particles to gradually shift from focusing on individual exploration to focusing on group cooperation during the search process, thereby improving the algorithm's convergence performance and search efficiency.

[0047] In one optional implementation, the exponential coefficient of the adaptive reaching law is adjusted according to the dynamic change of the deviation, while the ambient temperature change rate is collected and calculated in real time. The switching function gain coefficient of the sliding mode variable structure control algorithm is adaptively adjusted according to the ambient temperature change rate to form a composite adaptive control law, thereby achieving dynamic adjustment of the fan speed. This includes: The deviation change rate is calculated based on the deviation, and the sliding surface of the sliding mode variable structure control algorithm is constructed based on the deviation and the deviation change rate. The sliding surface is the product of the deviation and a preset constant plus the deviation change rate. The absolute value of the deviation and the absolute value of the deviation change rate are weighted and summed to obtain a comprehensive dynamic characteristic quantity. Based on the comprehensive dynamic characteristic quantity, the exponential coefficient of the adaptive reaching law is constructed. The exponential coefficient of the adaptive reaching law is formed by adding the basic exponential value to a negative exponential function with the comprehensive dynamic characteristic quantity as the independent variable. The ambient temperature is differentially calculated to obtain the rate of temperature change, and the switching function gain coefficient is designed based on the absolute value of the rate of temperature change. The switching function gain coefficient is a linear combination of the base gain value and the absolute value of the rate of temperature change. A first control term is constructed based on the exponential coefficients of the adaptive reaching law. The first control term includes the product of a variable exponential term of the absolute value of the sliding surface and a sign function, wherein the exponent of the variable exponential term is determined by the exponential coefficients of the adaptive reaching law. A second control term is constructed based on the gain coefficient of the switching function, and the second control term is the product of the gain coefficient of the switching function and the sign function of the sliding surface; The first control term and the second control term are superimposed to obtain a composite adaptive control law, and the composite adaptive control law is output as a control command to the fan execution system to realize the dynamic adjustment of the fan speed.

[0048] This embodiment provides a method for adjusting the coefficient of the adaptive approach law exponent based on the dynamic change of the deviation. This method can collect and calculate the rate of change of ambient temperature in real time, and adaptively adjust the gain coefficient of the switching function of the sliding mode variable structure control algorithm according to the rate of change of ambient temperature to form a composite adaptive control law, thereby realizing the dynamic adjustment of the fan speed.

[0049] In the process of wind turbine speed control, the actual value and set value of the wind turbine speed are first obtained, and the deviation between the two is calculated. To construct the sliding surface of the sliding mode variable structure control algorithm, the deviation change rate needs to be calculated based on the deviation. Specifically, the deviation change rate can be calculated using a difference method, which is the deviation value of the current sampling period minus the deviation value of the previous sampling period, and then divided by the sampling period time. In practical applications, assuming the deviation of the current sampling period is 2.5 rpm, the deviation of the previous sampling period is 2.8 rpm, and the sampling period is 0.1 seconds, then the deviation change rate is (2.5-2.8) / 0.1 = -3 rpm. 2 .

[0050] The sliding surface of the sliding mode variable structure control algorithm is constructed based on the calculated deviation and the rate of change of deviation. The sliding surface is defined as the product of the deviation and a preset constant plus the rate of change of deviation. In specific implementation, the preset constant can be set according to the system response characteristics, for example, 5.0. Continuing the previous example, the value of the sliding surface is 5.0 × 2.5 + (-3) = 9.5.

[0051] To achieve adaptive control, the absolute value of the deviation and the absolute value of the rate of change of deviation need to be weighted and summed to obtain a comprehensive dynamic characteristic. The weighting coefficients can be determined based on the system's sensitivity to the deviation and the rate of change of deviation; for example, the weight of the absolute value of the deviation is 0.7, and the weight of the absolute value of the rate of change of deviation is 0.3. Based on the example above, the comprehensive dynamic characteristic is 0.7 × |2.5| + 0.3 × |3| = 2.65.

[0052] The exponential coefficient of the adaptive reaching law is constructed based on the comprehensive dynamic characteristic. The exponential coefficient is formed by adding the basic exponential value to a negative exponential function with the comprehensive dynamic characteristic as the independent variable. The basic exponential value can be set to 0.5, and the negative exponential function can be expressed as exp(-comprehensive dynamic characteristic). Substituting into the example above, the exponential coefficient of the adaptive reaching law is 0.5 + exp(-2.65) ≈ 0.571.

[0053] In terms of ambient temperature control, the ambient temperature needs to be differentially calculated to obtain the temperature change rate. Assuming the current ambient temperature is 25.6℃, the temperature in the previous sampling period was 25.2℃, and the sampling period is 10 seconds, then the temperature change rate is (25.6-25.2) / 10=0.04℃ / second.

[0054] The switching function gain coefficient is designed based on the absolute value of the rate of temperature change. The switching function gain coefficient is a linear combination of the base gain value and the absolute value of the rate of temperature change. The base gain value can be set to 2.0, and the linear combination coefficient can be set to 25.0. Based on the example above, the switching function gain coefficient is 2.0 + 25.0 × |0.04| = 3.0.

[0055] The first control term is constructed based on the coefficients of the exponential term in the adaptive reaching law. The first control term consists of the product of a variable exponential term representing the absolute value of the sliding surface and a sign function. The exponent of the variable exponential term is determined by the coefficients of the exponential term in the adaptive reaching law. The value of the sign function depends on the sign of the sliding surface: 1 for a positive sign, -1 for a negative sign, and 0 for a zero sign. In the example above, the sliding surface value is 9.5, the sign function value is 1, and the first control term is |9.5|. 0.571 ×1≈3.1.

[0056] A second control term is constructed based on the gain coefficient of the switching function. The second control term is the product of the gain coefficient of the switching function and the sign function of the sliding surface. In the example above, the second control term is 3.0 × 1 = 3.0.

[0057] The composite adaptive control law is obtained by superimposing the first and second control terms. In the example above, the composite adaptive control law is 3.1 + 3.0 = 6.1. This control law is output as a control command to the fan actuator system to achieve dynamic adjustment of the fan speed.

[0058] In practical applications, this method demonstrates significant effectiveness in suppressing sudden disturbances. For example, when the ambient temperature suddenly rises, causing abrupt changes in deviation, the comprehensive dynamic characteristic quantity increases rapidly, leading to a decrease in the exponential coefficient of the adaptive reaching law, thus enhancing the system's robustness. Simultaneously, the increased rate of temperature change leads to an increase in the switching function gain coefficient, further strengthening the system's anti-interference capability.

[0059] This composite adaptive control method effectively addresses environmental temperature changes and load disturbances in the wind turbine speed control system, improving system response speed and stability while maintaining control accuracy. Experiments show that compared to traditional PID control, this method reduces overshoot by approximately 40%, steady-state error by approximately 60%, and dynamic response time by approximately 35% in wind turbine speed control.

[0060] In one optional implementation, after achieving dynamic adjustment of the fan speed, the method further includes: The inlet temperature in the real-time operating parameters is compared with the preset temperature threshold. When the inlet temperature is detected to be greater than the preset temperature threshold, the current real-time operating parameters are input into the hydraulic cooler heat dissipation power prediction model to calculate the target flow rate that can reduce the inlet temperature to below the preset temperature threshold. The pressure compensation correction value of the hydraulic system is calculated based on the target flow rate. The pressure compensation correction value is then superimposed on the current pressure compensation value of the hydraulic system. Through closed-loop control of pressure compensation, the hydraulic flow rate is gradually adjusted to the target flow rate, thereby achieving optimized control of the heat dissipation power of the hydraulic cooler.

[0061] To achieve dynamic adjustment of the fan speed and optimized control of the hydraulic cooler's heat dissipation power, the system continuously monitors the real-time operating parameters of the hydraulic system, especially the inlet temperature, and compares them with preset temperature thresholds. When the system detects that the inlet temperature exceeds the preset temperature threshold, it will initiate a series of optimization control measures.

[0062] The system first inputs the currently collected real-time operating parameters into the hydraulic cooler heat dissipation power prediction model. This prediction model, based on historical operating data and thermodynamic principles, calculates the required target flow rate based on the input parameters to reduce the inlet temperature below a preset temperature threshold. Real-time operating parameters include, but are not limited to, inlet temperature, ambient temperature, fan speed, hydraulic flow rate, and system pressure. For example, when the detected inlet temperature reaches 65℃, while the preset temperature threshold is 60℃, the system inputs the current ambient temperature (30℃), fan speed (1200rpm), hydraulic flow rate (80L / min), and system pressure (18MPa) into the prediction model.

[0063] The hydraulic cooler heat dissipation power prediction model calculates the target flow rate required to reduce the inlet temperature to below 60°C by analyzing the relationship between these parameters. This model considers the nonlinear relationship between hydraulic oil flow rate and heat dissipation power, and incorporates current environmental conditions and system status for prediction. In the example above, the prediction model calculates a target flow rate of 95 L / min, meaning that increasing the hydraulic flow rate from the current 80 L / min to 95 L / min will reduce the inlet temperature below the preset threshold.

[0064] After obtaining the target flow rate, the system calculates the required pressure compensation correction value for the hydraulic system. Pressure compensation is a crucial mechanism for controlling flow rate in a hydraulic system; adjusting the pressure compensation value effectively controls the output flow rate of the hydraulic pump. The calculation of the pressure compensation correction value considers the difference between the target flow rate and the current flow rate, as well as the characteristic parameters of the hydraulic system. For the example above, the system can calculate that a pressure compensation correction value of 1.5 MPa is needed to increase the flow rate from 80 L / min to 95 L / min.

[0065] The system adds the calculated pressure compensation correction value to the current pressure compensation value of the hydraulic system. Assuming the current pressure compensation value is 12 MPa, after adding a correction value of 1.5 MPa, the new pressure compensation setting value is 13.5 MPa. The system does not directly adjust the pressure compensation value to 13.5 MPa all at once; instead, it gradually adjusts the pressure compensation value through closed-loop control, allowing the hydraulic flow rate to smoothly increase to the target flow rate.

[0066] During closed-loop control, the system gradually adjusts the pressure compensation value in small steps. After each adjustment, it monitors the actual flow rate and adjusts the compensation value accordingly. For example, the system can first adjust the pressure compensation value to 12.5 MPa, monitoring an increase in flow rate to 85 L / min; then adjust it to 13 MPa, increasing the flow rate to 90 L / min; finally, adjust it to 13.5 MPa, achieving the target flow rate of 95 L / min. This gradual adjustment avoids drastic fluctuations in system pressure and flow, protecting hydraulic components and extending their service life.

[0067] Once the flow rate reaches the target value, the system continuously monitors changes in the inlet temperature. Due to the increased hydraulic flow, the heat dissipation power of the hydraulic cooler increases accordingly, and the inlet temperature begins to decrease. In the example case, the temperature can gradually decrease from 65°C to 58°C, below the preset temperature threshold of 60°C, at which point the goal of optimized heat dissipation power control has been achieved.

[0068] The system will not immediately reduce the hydraulic flow rate, but will continue to maintain it for a period of time to ensure that the temperature stabilizes below the threshold. For example, the system can maintain a flow rate of 95 L / min for about 5 minutes after the temperature drops to 58°C, and observe whether the temperature rises again. If the temperature remains stable or continues to drop, the system may appropriately reduce the flow rate to save energy, but will still ensure that the temperature does not exceed the preset threshold.

[0069] Throughout the optimization control process, the system also considers other constraints, such as maximum allowable flow rate and maximum system pressure. If the calculated target flow rate exceeds the system's maximum allowable flow rate, the system will limit the target flow rate within the maximum allowable value and trigger additional cooling measures, such as further increasing the fan speed.

[0070] In addition, the system dynamically adjusts the preset temperature threshold based on the usage of the hydraulic system. For example, under high load conditions, a slightly higher temperature threshold can be allowed; while under conditions requiring high precision control, a lower temperature threshold can be set to ensure system performance.

[0071] The above methods can optimize the heat dissipation power of the hydraulic cooler, ensuring that the hydraulic system maintains a suitable operating temperature under various working conditions, extending equipment life and improving system efficiency.

[0072] The hydraulic cooler heat dissipation power optimization system based on airflow according to an embodiment of the present invention includes: The first unit is used to acquire the real-time operating parameters of the hydraulic cooler, calculate the current heat dissipation power of the hydraulic cooler based on the real-time operating parameters, and use the current heat dissipation power as the actual heat dissipation power of the hydraulic cooler. The second unit is used to input the real-time operating parameters into the hydraulic cooler heat dissipation power prediction model, and calculate the target fan speed that maximizes the predicted heat dissipation power based on the improved particle swarm optimization algorithm. The improved particle swarm optimization algorithm introduces an adaptive adjustment strategy for inertial weights and a dynamic convergence factor on the basis of the original particle swarm optimization algorithm. When the power error between the predicted heat dissipation power of the hydraulic cooler heat dissipation power prediction model and the actual heat dissipation power is greater than a preset power threshold, an optimization objective function is constructed based on a deep reinforcement learning algorithm. The power error is used as the reward function, and the network parameters of the hydraulic cooler heat dissipation power prediction model are iteratively optimized online using the policy gradient method until the power error is less than the preset power threshold. The third unit is used to establish a closed-loop control system for wind turbine speed based on the sliding mode variable structure control algorithm. It constructs a sliding mode switching surface based on the deviation between the target wind turbine speed and the current wind turbine speed, and uses an adaptive approach law to achieve rapid tracking control of the wind turbine speed. Specifically, the exponential coefficient of the adaptive approach law is adjusted according to the dynamic changes in the deviation, while the ambient temperature change rate is collected and calculated in real time. The gain coefficient of the switching function of the sliding mode variable structure control algorithm is adaptively adjusted according to the ambient temperature change rate to form a composite adaptive control law, thereby achieving dynamic adjustment of the wind turbine speed.

[0073] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0074] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0075] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing the heat dissipation power of a hydraulic cooler based on airflow, characterized in that, include: The real-time operating parameters of the hydraulic cooler are obtained, the current heat dissipation power of the hydraulic cooler is calculated based on the real-time operating parameters, and the current heat dissipation power is used as the actual heat dissipation power of the hydraulic cooler. The real-time operating parameters are input into the hydraulic cooler heat dissipation power prediction model. The target fan speed that maximizes the predicted heat dissipation power is calculated based on the improved particle swarm optimization algorithm. The improved particle swarm optimization algorithm introduces an adaptive adjustment strategy for inertial weights and a dynamic convergence factor on the basis of the original particle swarm optimization algorithm. When the power error between the predicted heat dissipation power of the hydraulic cooler heat dissipation power prediction model and the actual heat dissipation power is greater than a preset power threshold, an optimization objective function is constructed based on a deep reinforcement learning algorithm. The power error is used as the reward function, and the network parameters of the hydraulic cooler heat dissipation power prediction model are iteratively optimized online using the policy gradient method until the power error is less than the preset power threshold. A closed-loop control system for wind turbine speed is established based on a sliding mode variable structure control algorithm. A sliding mode switching surface is constructed according to the deviation between the target wind turbine speed and the current wind turbine speed. An adaptive reaching law is used to achieve rapid tracking control of the wind turbine speed. Specifically, the exponential coefficient of the adaptive reaching law is adjusted according to the dynamic change of the deviation. At the same time, the ambient temperature change rate is collected and calculated in real time. The gain coefficient of the switching function of the sliding mode variable structure control algorithm is adaptively adjusted according to the ambient temperature change rate to form a composite adaptive control law, thereby realizing the dynamic adjustment of the wind turbine speed.

2. The method according to claim 1, characterized in that, An optimization objective function is constructed based on a deep reinforcement learning algorithm, and the power error is used as the reward function. The network parameters of the hydraulic cooler heat dissipation power prediction model are then iteratively optimized online using a policy gradient method, including: A multi-objective reward function is constructed, which includes three weighted product terms. The first weighted product term is the product of the first weight coefficient and the exponential decay value of the power error. The second weighted product term is the product of the second weight coefficient and the exponential decay value of the power error change rate. The third weighted product term is the product of the third weight coefficient and the prediction smoothness evaluation value. The policy gradient optimization objective is constructed based on the multi-objective reward function. The expected value of the product of the log-likelihood value of the policy network parameters and the discount reward is used as the policy gradient calculation formula. The policy gradient value is calculated by the policy gradient calculation formula, wherein the discount reward is calculated by the multi-objective reward function. The dynamic learning rate is calculated based on the value estimate of the current state, where the value estimate represents the expected cumulative reward in the current state, and the dynamic learning rate is the product of the initial learning rate and the exponential decay function of the value estimate. The network parameter adjustment amount is obtained by multiplying the policy gradient value by the dynamic learning rate. The network parameter adjustment amount is then used to update the current network parameters to obtain the updated network parameters. The magnitude of the network parameter adjustment amount is adaptively adjusted according to the value estimate. The updated network parameters are subjected to anomaly detection, and the final network parameters are determined using a parameter rollback mechanism and updated in the hydraulic cooler heat dissipation power prediction model.

3. The method according to claim 1, characterized in that, The target fan speed that maximizes the predicted heat dissipation power is calculated based on an improved particle swarm optimization (PSO) algorithm. This improved PSO algorithm introduces an adaptive inertia weight adjustment strategy and a dynamic convergence factor, based on the original PSO algorithm, including: An improved particle swarm optimization algorithm is constructed by adaptively adjusting the inertia weight. The difference between the maximum and minimum inertia weights is multiplied by an exponential decay function based on the current iteration number. The product is then subtracted from the maximum inertia weight to obtain the inertia weight for the current iteration. The improved particle swarm optimization algorithm is constructed using a dynamic convergence factor, which includes a cognitive learning factor for adjusting the influence of the individual particle's historical best position on the particle's motion, and a social learning factor for adjusting the influence of the swarm's historical best position on the particle's motion. The product of the inertial weight and the particle's current velocity, the product of the cognitive learning factor and the individual optimal position deviation, the product of the social learning factor and the global optimal position deviation, and the weighted sum of the flight disturbance term are used as the particle's velocity update. Calculate a population diversity index, construct an adaptive mutation probability using the population diversity index, wherein the adaptive mutation probability decreases exponentially as the population diversity index increases, and perform mutation operations on the particle swarm based on the adaptive mutation probability. The mutated particle positions are substituted into the heat dissipation power prediction model to obtain the corresponding predicted heat dissipation power. Based on the predicted heat dissipation power, fitness evaluation is performed. The particle position with the best evaluation result is saved to the elite solution set, and the elite solution set is dynamically updated in each iteration. The wind turbine speed corresponding to the position of the particle with the best fitness in the elite solution set is taken as the target wind turbine speed.

4. The method according to claim 3, characterized in that, The method further includes: The cognitive learning factor is calculated by adding the product of the first random initial value and the square of the ratio of the current iteration number to the maximum iteration number; The social learning factor is calculated by adding the product of the second random initial value and the square of the ratio of the current iteration number to the maximum iteration number; The individual optimal position deviation is the difference between the individual particle's historical optimal position and its current position, while the global optimal position deviation is the difference between the collective particle population's historical optimal position and its current position.

5. The method according to claim 1, characterized in that, The exponential coefficient of the adaptive reaching law is adjusted according to the dynamic change of the deviation. Simultaneously, the rate of change of ambient temperature is collected and calculated in real time. The gain coefficient of the switching function of the sliding mode variable structure control algorithm is adaptively adjusted according to the rate of change of ambient temperature, forming a composite adaptive control law to achieve dynamic adjustment of the fan speed, including: The deviation change rate is calculated based on the deviation, and the sliding surface of the sliding mode variable structure control algorithm is constructed based on the deviation and the deviation change rate. The sliding surface is the product of the deviation and a preset constant plus the deviation change rate. The absolute value of the deviation and the absolute value of the deviation change rate are weighted and summed to obtain a comprehensive dynamic characteristic quantity. Based on the comprehensive dynamic characteristic quantity, the exponential coefficient of the adaptive reaching law is constructed. The exponential coefficient of the adaptive reaching law is formed by adding the basic exponential value to a negative exponential function with the comprehensive dynamic characteristic quantity as the independent variable. The ambient temperature is differentially calculated to obtain the rate of temperature change, and the switching function gain coefficient is designed based on the absolute value of the rate of temperature change. The switching function gain coefficient is a linear combination of the base gain value and the absolute value of the rate of temperature change. A first control term is constructed based on the exponential coefficients of the adaptive reaching law. The first control term includes the product of a variable exponential term of the absolute value of the sliding surface and a sign function, wherein the exponent of the variable exponential term is determined by the exponential coefficients of the adaptive reaching law. A second control term is constructed based on the gain coefficient of the switching function, and the second control term is the product of the gain coefficient of the switching function and the sign function of the sliding surface; The first control term and the second control term are superimposed to obtain a composite adaptive control law, and the composite adaptive control law is output as a control command to the fan execution system to realize the dynamic adjustment of the fan speed.

6. The method according to claim 1, characterized in that, After achieving dynamic adjustment of the fan speed, the method further includes: The inlet temperature in the real-time operating parameters is compared with the preset temperature threshold. When the inlet temperature is detected to be greater than the preset temperature threshold, the current real-time operating parameters are input into the hydraulic cooler heat dissipation power prediction model to calculate the target flow rate that can reduce the inlet temperature to below the preset temperature threshold. The pressure compensation correction value of the hydraulic system is calculated based on the target flow rate. The pressure compensation correction value is then superimposed on the current pressure compensation value of the hydraulic system. Through closed-loop control of pressure compensation, the hydraulic flow rate is gradually adjusted to the target flow rate, thereby achieving optimized control of the heat dissipation power of the hydraulic cooler.

7. A hydraulic cooler heat dissipation power optimization system for airflow, used to implement the method as described in any one of claims 1-6, characterized in that, include: The first unit is used to acquire the real-time operating parameters of the hydraulic cooler, calculate the current heat dissipation power of the hydraulic cooler based on the real-time operating parameters, and use the current heat dissipation power as the actual heat dissipation power of the hydraulic cooler. The second unit is used to input the real-time operating parameters into the hydraulic cooler heat dissipation power prediction model, and calculate the target fan speed that maximizes the predicted heat dissipation power based on the improved particle swarm optimization algorithm. The improved particle swarm optimization algorithm introduces an adaptive adjustment strategy for inertial weights and a dynamic convergence factor on the basis of the original particle swarm optimization algorithm. When the power error between the predicted heat dissipation power of the hydraulic cooler heat dissipation power prediction model and the actual heat dissipation power is greater than a preset power threshold, an optimization objective function is constructed based on a deep reinforcement learning algorithm. The power error is used as the reward function, and the network parameters of the hydraulic cooler heat dissipation power prediction model are iteratively optimized online using the policy gradient method until the power error is less than the preset power threshold. The third unit is used to establish a closed-loop control system for wind turbine speed based on the sliding mode variable structure control algorithm. It constructs a sliding mode switching surface based on the deviation between the target wind turbine speed and the current wind turbine speed, and uses an adaptive approach law to achieve rapid tracking control of the wind turbine speed. Specifically, the exponential coefficient of the adaptive approach law is adjusted according to the dynamic changes in the deviation, while the ambient temperature change rate is collected and calculated in real time. The gain coefficient of the switching function of the sliding mode variable structure control algorithm is adaptively adjusted according to the ambient temperature change rate to form a composite adaptive control law, thereby achieving dynamic adjustment of the wind turbine speed.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images