Evaporative condenser energy-saving optimization method based on DDPG-GA fusion

Abstract

The invention discloses an evaporative condenser energy-saving optimization method based on DDPG-GA fusion, and the method comprises the following steps: 1, obtaining historical data of a refrigeration system, carrying out the data preprocessing, removing abnormal values in the data, carrying out the standardization processing, and finally dividing a data set; step 2, using a deep neural network to construct a condensation pressure prediction model as a system simulator, and using a PCA method to obtain a dependency rule between parameters as a constraint condition of the model; the method has the beneficial effects that an energy-saving optimization control strategy based on fusion of deep reinforcement learning (DDPG) and a genetic algorithm (GA) is introduced, a deep neural network condensation pressure prediction model is established as a system simulation environment, and a reward function is designed, so that the control strategy can automatically adjust parameters according to real-time working conditions; therefore, complex working conditions are dynamically adapted, and global optimization of energy efficiency is realized.

Description

Technical Field

[0001] This invention relates to the field of condenser technology, and specifically to an energy-saving optimization method for evaporative condensers based on DDPG-GA fusion. Background Technology

[0002] With the continuous development of industrial automation and intelligent manufacturing, refrigeration systems, as an important component of modern industrial and commercial buildings, have attracted much attention regarding their energy efficiency and operational stability. Evaporative condensers, as key equipment in refrigeration systems, directly affect the overall system performance and energy consumption through their operating status. However, traditional control methods struggle to dynamically adapt to rapid changes in the environment and load, resulting in unsatisfactory energy-saving effects and often failing to simultaneously meet multiple objectives such as response speed, steady-state accuracy, and energy consumption optimization.

[0003] In recent years, deep reinforcement learning (DRL) has attracted widespread attention as an advanced method for handling continuous control problems. Among them, the deep deterministic policy gradient algorithm (DDPG) has been applied to the control of complex industrial systems due to its excellent performance in continuous action spaces. However, there are relatively few mature deep reinforcement learning energy-saving optimization strategies for refrigeration systems, and it also has limitations such as low environmental exploration degree, low data utilization, and sparse rewards. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide an energy-saving optimization method for evaporative condensers based on DDPG-GA fusion. By introducing deep learning and intelligent optimization technology, it provides a brand-new energy-saving control solution that can dynamically adapt to complex and ever-changing environmental conditions and achieve global energy efficiency optimization.

[0005] The technical solution adopted by this invention to solve its technical problem is: an energy-saving optimization method for evaporative condensers based on DDPG-GA fusion, comprising the following steps:

[0006] Step 1: Obtain historical data of the refrigeration system, perform data preprocessing, remove outliers, standardize the data, and finally divide the dataset.

[0007] Step 2: Use a deep neural network to build a condensation pressure prediction model as a system simulator, and use the PCA method to obtain the dependency rules between parameters as constraints of the model, so that every action in the deep reinforcement learning process conforms to the actual working conditions and solves the sparse reward problem.

[0008] Step 3: Establish an energy-saving optimization strategy model for evaporative condensers based on the improved DDPG. Use the trained pressure prediction model as the environment for deep reinforcement learning, construct the Actor network and Critic network, design a dual experience pool mechanism, design a reward function, and use the deviation between the condensing pressure and the target pressure as the core index to find the system parameter set that minimizes the pressure error.

[0009] Step 4: Integrate genetic algorithms into the DDPG reinforcement learning framework for efficient exploration: Generate an initial population based on the current Actor network strategy, evaluate the fitness of individuals in the population through a simulator, then perform selection, crossover, and mutation operations to optimize the population, and finally add the best optimized individuals to the experience pool.

[0010] Step 5, Online Training and Optimized Control: After training is completed, the trained Actor network is deployed to the actual control system to receive real-time status inputs and output control actions; at the same time, the system can continue to collect actual operating data, update the model periodically, and achieve continuous optimization.

[0011] Furthermore, in step 1, the preprocessing step includes removing outliers from the historically collected operating parameters and environmental data. After removal, mean imputation is used to fill in missing values. StandardScaler is used to standardize the input features and output target, converting each feature into a distribution with zero mean and unit standard deviation to eliminate the influence of dimensions. The preprocessed data is then integrated into a sample set and divided into training and testing datasets in an 8:2 ratio. Z-Score standardization is applied to each feature dimension, and the specific transformation function is as follows:

[0012]

[0013] In the formula, μ is the mean of the input feature across all samples, and σ is the standard deviation of the feature across all samples.

[0014] Furthermore, in step 2, a three-layer feedforward fully connected neural network, consisting of one input layer, three hidden layers, and one output layer, is selected to establish the system model. Considering the characteristics of the input and output parameters of the neural network model in this method, the ReLU function is chosen as the activation function for the hidden layers, and its formula is:

[0015] ReLU = max(0, z) (2)

[0016] The neural network is trained using the Adam optimizer, with the loss function being the mean squared error (MSE), and its formula is as follows:

[0017]

[0018] In the formula, y iIt is the true value of the i-th sample. It is the model's predicted value for the i-th sample, and n is the total number of samples.

[0019] Furthermore, implicit dependencies between parameters are extracted using principal component analysis (PCA), and these extracted dependencies are applied as constraints to the action space of the DDPG algorithm. The core of this approach lies in calculating the data covariance matrix C, with the following formula:

[0020]

[0021] In the formula, N is the number of samples, μ is the mean vector of the training data, and x i These are the standardized input features.

[0022] Furthermore, in step 3, the goal of the Actor network is to generate actions that maximize the expected reward for the environment. To achieve this, its parameters need to be updated using gradient ascent, ensuring that the action output in the current state achieves a higher Q-value under the evaluation of the Critic network. The gradient update formula for the Actor network is:

[0023]

[0024] In the formula, θ μ These are the parameters of the Actor network, μ(s|θ) μ ) is an Actor network, Q(s,a|θ) Q ) is the Critic network, and N is the batch size.

[0025] The task of the Critic network is to accurately estimate the Q-value of the state-action pair (s, a). Its updates are based on temporal difference (TD) learning, achieved by minimizing the mean squared error (MSE) between the current predicted Q-value and the target Q-value. The MSE loss function of the Critic network is:

[0026]

[0027] The target Q-value is calculated through the target network using the following formula:

[0028] y i =r i +γQ′(s i+1 ,μ′(s i+1 |θ μ′ )|θ Q′ (19)

[0029] In the formula, θ Q These are the parameters of the Critic network, θ Q′ and θ μ′ These are the target network parameters, ri It's an instant reward, γ is the discount sub-Q. ′ and μ ′ These are the target Critic and Actor networks, respectively.

[0030] To further improve the update performance of the Critic network, a weighting mechanism can be introduced when calculating the loss. The loss of each sample in the batch is summed using weighted averages. The weighted loss function formula for Critic is as follows:

[0031]

[0032] In the formula, N is the number of samples in the batch, and D is the dimension of the Critic output (if the output is a scalar, then D = 1); Q(s i ,a i ) d It is the Q-value calculated by the current Critic network in the d-th dimension for the i-th sample; y i,d This corresponds to the TD target value, with weight w. i The importance of sampling is calculated based on TD error.

[0033] Furthermore, in step 3, the DDPG algorithm employs a dual experience pool mechanism, setting up a new experience pool and an old experience pool. The new experience pool stores all experience samples collected in real-time from the interaction with the environment during the current training round, while the old experience pool stores experience samples from historical training rounds whose reward values ​​are higher than the average reward value. During training, experience samples are sampled from both the new and old experience pools at a preset ratio to update the Critic network. The formula for selecting high-reward samples in the old experience pool is:

[0034]

[0035] In the formula, r t is the immediate reward value for the t-th experience sample, and M is the total number of samples currently stored in the old experience pool. It sums up the reward values ​​of all samples in the old experience pool.

[0036] Furthermore, in step 4, reinforcement learning actors are added to the population. Each actor in the population interacts with the environment, undergoing crossover and mutation to generate a new population. Experiences with high fitness are added to the experience pool for DDPG to learn from. The fitness calculation formula is as follows:

[0037] Fitness (a i )=Q φ (s,a i )+α·r(s,a i (26)

[0038] In the formula, Qφ (s,a i ): Critic network for action a i Long-term return estimate, r(s,a) i Action a i The immediate reward, α: a weighting coefficient that balances long-term and short-term returns.

[0039] Furthermore, in step 4, to balance the policy utilization of deep reinforcement learning with the global exploration capability of evolutionary algorithms, the initial population of the genetic algorithm is constructed by applying Gaussian perturbation after generating baseline action vectors through the current Actor network. The specific formula is as follows:

[0040]

[0041] In the formula, i = 1, 2, ..., N represents candidate individuals for actions within the population, σ is the standard deviation of the perturbation, and clip(a0+∈ i ,-a max ,a max Ensure that the actions are within the permissible range.

[0042] Furthermore, in step 4, the mutation operation employs adaptive Gaussian noise. In the DDPG framework, the policy update of the Actor network is affected by locally optimal actions. While generating diverse candidate actions through genetic algorithms can expand the exploration range, traditional random mutation may produce a large number of invalid actions, reducing learning efficiency. This mechanism can effectively alleviate the sparse reward and local convergence problems of reinforcement learning. The formula for calculating the mutation intensity σ is:

[0043]

[0044] In the formula, σ0 represents the basic standard deviation of variation, k represents the adjustment coefficient, and σ fitness denoted as the standard deviation of fitness of each individual in the current population, ∈ represents a small constant to prevent division by zero.

[0045] The beneficial effects of this invention are achieved by introducing an energy-saving optimization control strategy based on the fusion of deep reinforcement learning (DDPG) and genetic algorithm (GA), establishing a deep neural network condensing pressure prediction model as the system simulation environment, and designing a reward function to enable the control strategy to automatically adjust parameters according to real-time operating conditions, thereby dynamically adapting to complex operating conditions and achieving global optimization of energy efficiency. Furthermore, addressing the problems of low environment exploration, low data utilization, and sparse rewards inherent in deep reinforcement learning, this invention employs PCA projection constraints and a dual-experience pool mechanism to improve sample utilization efficiency, and introduces a genetic algorithm into the DDPG framework for diversified population evolution optimization to enhance environment exploration capabilities. Simultaneously, PCA projection ensures that each control action conforms to the actual data distribution, thereby alleviating the sparse reward problem. In summary, the technical solution adopted in this invention can significantly improve the regulation performance and energy efficiency of evaporative condenser systems, balancing response speed, control accuracy, and energy efficiency optimization, achieving significant technological progress and comprehensive performance improvement compared to existing technologies. Attached Figure Description

[0046] Figure 1 This is a basic flowchart of an energy-saving optimization method for evaporative condensers based on DDPG-GA fusion.

[0047] Figure 2 This is a schematic diagram of an energy-saving optimization method for evaporative condensers based on DDPG-GA fusion.

[0048] Figure 3 It is based on the convergence speed of the DDPG method.

[0049] Figure 4 It is based on the convergence speed of the DDPG-GA method. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Figure 1 and Figure 2 As shown, this invention provides an energy-saving optimization method for evaporative condensers based on DDPG-GA fusion, the structure of which is as follows. Figure 2 As shown, firstly, a condensing pressure prediction model is constructed using data preprocessing and a deep neural network as a system simulator to provide an environment model for subsequent reinforcement learning. Then, based on the constructed simulator, a reinforcement learning environment is designed, including the definitions of state, action, and reward. Next, the DDPG-GA fusion algorithm module is responsible for training the agent to learn the optimal control strategy. Finally, the online training and optimization control module applies the trained strategy to the actual system control and continuously optimizes the strategy based on real-time feedback. The specific steps are as follows:

[0051] Step 1: Obtain historical data of the evaporative condenser and perform data preprocessing. Delete data that deviates significantly from the normal range to remove outliers. Standardize each feature dimension using Z-Score. Finally, divide the dataset.

[0052] Step 1-1: Use the Z-Score strategy (1) for outlier handling. The transformation function is as follows:

[0053]

[0054] In the formula, μ is the mean of the input feature across all samples, and σ is the standard deviation of the feature across all samples.

[0055] Steps 1-2: Divide the preprocessed data into training and testing datasets in an 8:2 ratio.

[0056] Step 2: Construct a condensation pressure prediction model using a deep neural network. Through multi-layer linear transformation and ReLU activation function, the standardized input operating parameters are mapped to the predicted pressure, and the mean squared error is used as the loss function for training. At the same time, PCA is used to reduce the dimensionality of the training data, extract the main feature directions, and obtain the principal component matrix. Then, the actions generated by the DDPG agent are projected into the main subspace of the data distribution, thereby ensuring that each action conforms to the actual operating conditions and alleviating the problem of sparse rewards in the reinforcement learning process.

[0057] Step 2-1: The neural network for the condensation pressure prediction model adopts a multi-layer feedforward neural network structure, specifically consisting of one input layer, three hidden layers, and one output layer. The ReLU function is selected as the activation function for the hidden layers, and its formula is as follows:

[0058] ReLU = max(0, z) (2)

[0059] The forward propagation formula for this deep neural network is as follows:

[0060] h1=ReLU(W1x+b1) (3)

[0061] h2=ReLU(W2x+b2) (4)

[0062] h3=ReLU(W3x+b3) (5)

[0063]

[0064] h i =ReLU(W i x+b i (7)

[0065] In the formula, x represents the input feature vector (after standardization). W i Let represent the weight matrix of the i-th layer, used for linear transformation. b i `z` represents the bias vector of the i-th layer. `ReLU` represents the activation function, with the form `ReLU(z) = max(0,z)`, which increases the non-linear expressive power of the network. i This represents the output (hidden layer feature) of the i-th layer. This represents the final output of the neural network.

[0066] The loss function uses mean squared error (MSE), and its expression is as follows:

[0067]

[0068] In the formula, y i It is the true value of the i-th sample. It is the model's predicted value for the i-th sample, and n is the total number of samples.

[0069] Step 2-2: To ensure that each action generated by the DDPG agent conforms to actual working conditions and to alleviate the sparse reward problem caused by state jumps in reinforcement learning, PCA projection constraints are applied to the actions. First, data dimensionality reduction and feature extraction are performed. After standardizing the training data, the covariance matrix is ​​calculated using the following formula:

[0070]

[0071] In the formula, N is the number of samples, μ is the mean vector of the training data, and x i These are the standardized input features.

[0072] The eigenvalue decomposition formula is as follows:

[0073] C = QΛQ T (10)

[0074] In the formula, Q represents the eigenvector matrix, and Λ represents the diagonal matrix.

[0075] Next, principal components are selected, with the top k principal components chosen to achieve a cumulative contribution rate of 95%.

[0076]

[0077] In the formula, λ i Let represent the i-th largest eigenvalue, k represent the number of principal components selected such that the cumulative contribution rate of the first k features reaches 95%, and d represent the dimension of the original features.

[0078] To ensure that the state remains within the main region of the actual data distribution after each action update and to prevent excessive state shifts, the original action vector 'a' generated by the DDPG agent (directly output to the Actor network) is projected into the principal component subset space. The projection operation formula is:

[0079] a proj =V(V T a) (12)

[0080] In the formula, a represents the original action vector (output by the reinforcement learning agent), V represents the principal component matrix obtained by PCA, and V T Denotes the transpose of V; a proj This indicates the action after projection, ensuring that the direction of change lies in the principal components of the data.

[0081] By using PCA projection as a soft constraint on actions, the exploration direction is guided in the early stages of training. When the Critic network evaluates that a new action significantly improves the Q value, it is allowed to deviate from the historical distribution to a certain extent, thereby alleviating the sparse reward problem caused by discontinuous states or deviations from actual working conditions during reinforcement learning.

[0082] Step 3: Based on the constructed neural network simulator, design the reinforcement learning environment: Construct a policy network (Actor) and a value network (Critic) to learn the control policy and state-action value, respectively. Update network parameters using TD error and policy gradient, with condensing pressure deviation as the core optimization index. Design a reward function to drive policy learning. Introduce a dual experience pool mechanism and importance sampling technique to improve the sample efficiency and stability of the DDPG algorithm.

[0083] Step 3-1: In the reinforcement learning environment, the trained condensing pressure prediction model is first used as the system simulator to calculate the condensing pressure at the next time step. Given a state s, the data is standardized and then input into the prediction model. The model outputs the standardized condensing pressure, which is then de-standardized to obtain the predicted pressure.

[0084]

[0085] The condensation pressure and the target pressure p target The reward function is defined with the absolute error as the core, and the formula is as follows:

[0086]

[0087] Step 3-2: Design the reinforcement learning environment and construct the policy network (Actor) and value network (Critic). The Actor network generates the control action 'a' from the current state 's', and its mathematical expression is:

[0088]

[0089] In the formula, s represents the current state (the standardized values ​​of the system parameters), and g θμ (s) represents the mapping function of the Actor network to state s, with parameters θ. μ , This represents the hyperbolic tangent activation function, which restricts the output to the range [-1, 1]. max This represents the maximum amplitude of the defined action, used to amplify the output after passing through tanh, so that the actual action range is [-a]. max ,a max ], where 'a' represents the final output action, used to adjust system parameters.

[0090] The gradient update formula for the Actor network is:

[0091]

[0092] In the formula, θ μ These are the parameters of the Actor network, μ(s|θ) μ ) is an Actor network, Q(s,a|θ) Q ) is the Critic network, and N is the batch size;

[0093] The Critic network evaluates the Q-value of a given state-action pair (s, a), representing the cumulative reward obtained by taking that action in that state. Its mathematical expression is:

[0094]

[0095] In the formula, This represents the mapping function of the Critic network for states s and actions a, with parameters θ. Q Q(s,a) represents the Q value corresponding to a given state and action, which represents the expected cumulative reward for taking the action in that state.

[0096] The goal of the Critic network is to minimize the mean squared error (MSE) between the predicted Q-value and the target Q-value. The MSE loss function of the Critic network is:

[0097]

[0098] In the formula, θ Q These are the parameters of the Critic network, where N is the batch size and y is the critic network parameter. i Denotes the target Q value, Q(s) i ,a i |θ Q ) is the Critic network's prediction of the state.

[0099] The target Q value is calculated as follows:

[0100] y i =r i +γQ′(s i+1 ,μ′(s i+1 |θ μ′ )|θ Q′ (19)

[0101] In the formula, θ Q′ and θ μ′ These are the target network parameters, r i It's an instant reward, γ is the discount sub-Q. ′ and μ ′ These are the target Critic and Actor networks, respectively.

[0102] The weighted loss function formula for Critic is:

[0103]

[0104] In the formula, N is the number of samples in the batch, and D is the dimension of the Critic output (if the output is a scalar, then D = 1); Q(s i ,a i ) d It is the Q-value calculated by the current Critic network in the d-th dimension for the i-th sample; y i,d This corresponds to the TD target value. Weight w i The formula for calculating the sampling importance based on TD error is as follows:

[0105]

[0106] In the formula, ∈ is a small constant to prevent the denominator from being zero, w min and w max These are the lower and upper limits of importance weight, respectively, used to prevent the weight from being too large or too small.

[0107] Step 3-3: Design two experience pools (a new experience pool and an old experience pool). The new experience pool stores the latest samples collected in the current round, and the old experience pool stores samples from previous rounds. During each training session, samples are drawn proportionally from the two experience pools to ensure a balance between new and old experiences. After each round, samples from the new experience pool are moved to the old experience pool. The formula for selecting high-reward samples in the old experience pool is:

[0108]

[0109] In the formula, r t is the immediate reward value for the t-th experience sample, and M is the total number of samples currently stored in the old experience pool. It sums up the reward values ​​of all samples in the old experience pool.

[0110] Step 4: Place the reinforcement learning action actors into the population. Each actor in the population interacts with the environment, and crossover and mutation occur to generate a new population. Experiences with high fitness are placed into the experience pool for DDPG to learn.

[0111] Step 4-1: Generate an initial population based on the current Actor. For the current state s, first generate basic actions through the current Actor network:

[0112] a0=A(s|θ A ) (twenty three)

[0113] Based on the basic action a0, a random perturbation is added to construct an initial population:

[0114]

[0115] In the formula, i = 1, 2, ..., N represents candidate individuals for actions within the population, σ is the standard deviation of the perturbation, and clip(a0+∈ i ,-a max ,a max Ensure that the actions are within the permissible range.

[0116] Step 4-2: Perform individual assessment and fitness calculation for each candidate action a in the population. i In the current state s, the environment simulator performs one step to obtain the next state s′. i With instant rewards i :

[0117] (s′ i ,r i ,d i ,...)=Env.step(a i (25)

[0118] Estimating the long-term payoff using a Critic network for the current state and candidate actions: Q(s,a) i |θ Q )

[0119] The fitness function is designed to comprehensively consider both long-term benefits and immediate rewards. A commonly used form is:

[0120] Fitness (a i )=Q φ (s,a i )+α·r(s,a i (26)

[0121] In the formula, Q φ (s,ai ) indicates that the Critic network is responsible for action a. i Long-term return estimate, r(s,a) i ) represents action a i The immediate reward, where α represents the weighting coefficient that balances long-term and short-term returns.

[0122] Step 4-3: For the two selected parent individuals θ parent1 and θ parent2 Offspring are generated using linear weighting:

[0123] θ child =αθ parent1 +(1-α)θ parent2 (27)

[0124] In the formula, θ parent1 ,θ parent2 This represents the parameters of the parent Actor. α represents the cross weight, which is typically sampled from a uniform distribution U(0,1).

[0125] Step 4-4: After crossover, mutate the offspring parameters using additive Gaussian noise.

[0126]

[0127] Wherein, the variation intensity σ is obtained through adaptive calculation, and its formula is:

[0128]

[0129] In the formula, σ0 represents the basic standard deviation of variation; j represents the adjustment coefficient, used to control the effect of fitness gap on the amplitude of variation; σ fitness This represents the standard deviation of the fitness of each individual in the current population. ∈ represents a small constant to prevent division by zero. This represents Gaussian noise with a mean of 0 and a standard deviation of σ.

[0130] The parent generation is merged with the offspring resulting from crossover mutation to form a new generation of population. After repeating the above steps for multiple iterations, the population gradually evolves to a better state.

[0131] Step 5: To verify the performance of this invention, a baseline comparison experiment was conducted using the DDPG method. DDPG is a deep deterministic policy gradient algorithm that can learn more effectively in a continuous high-dimensional action space and obtain optimization results faster with fewer samples in optimization problems. In this embodiment, DDPG is implemented using the PyTorch framework in Python.

[0132] like Figure 3 , Figure 4As shown, the GA-DDPG fusion algorithm exhibits a steeper early rise curve compared to the DDPG method, indicating that genetic exploration accelerates the initial learning. Therefore, the method proposed in this invention has the fastest convergence speed and a more stable convergence curve with less fluctuation.

[0133] 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. An energy-saving optimization method for evaporative condensers based on DDPG-GA fusion, characterized in that: Includes the following steps: Step 1: Obtain historical data of the refrigeration system, perform data preprocessing, remove outliers, standardize the data, and finally divide the dataset. Step 2: Use a deep neural network to build a condensation pressure prediction model as a system simulator, and use the PCA method to obtain the dependency rules between parameters as constraints for the model. Step 3: Establish an energy-saving optimization strategy model for evaporative condensers based on the improved DDPG. Use the trained pressure prediction model as the environment for deep reinforcement learning, construct the Actor network and Critic network, design a dual experience pool mechanism, design a reward function, and use the deviation between the condensing pressure and the target pressure as the core index to find the system parameter set that minimizes the pressure error. Step 4: Integrate genetic algorithms into the DDPG reinforcement learning framework for efficient exploration: Generate an initial population based on the current Actor network strategy, evaluate the fitness of individuals in the population through a simulator, then perform selection, crossover, and mutation operations to optimize the population, and finally add the best optimized individuals to the experience pool. Step 5: Online Training and Optimized Control: After training is completed, the trained Actor network is deployed to the actual control system to receive real-time status inputs and output control actions. At the same time, the system continues to collect actual operating data, updates the model regularly, and achieves continuous optimization.

2. The energy-saving optimization method for evaporative condensers based on DDPG-GA fusion according to claim 1, characterized in that: In step 1, the preprocessing step includes removing outliers from the historically collected operating parameters and environmental data. After removal, mean imputation is used to fill in missing values. StandardScaler is used to standardize the input features and output target, converting each feature into a distribution with zero mean and unit standard deviation to eliminate the influence of dimensions. The preprocessed data is then integrated into a sample set and divided into training and testing datasets in an 8:2 ratio. Z-Score standardization is applied to each feature dimension, and the transformation function is as follows: In the formula, μ is the mean of the input feature across all samples, and σ is the standard deviation of the feature across all samples.

3. The energy-saving optimization method for evaporative condensers based on DDPG-GA fusion according to claim 1, characterized in that: In step 2, a three-layer feedforward fully connected neural network, consisting of one input layer, three hidden layers, and one output layer, is selected to establish the system model. Considering the characteristics of the input and output parameters of the neural network model in this method, the ReLU function is chosen as the activation function for the hidden layers, and its formula is as follows: ReLU = max(0, z) (2) The neural network is trained using the Adam optimizer, with the loss function being the mean squared error (MSE), and its formula is as follows: In the formula, y i It is the true value of the i-th sample. It is the model's predicted value for the i-th sample, and n is the total number of samples.

4. The energy-saving optimization method for evaporative condensers based on DDPG-GA fusion according to claim 1, characterized in that: In step 3, the goal of the Actor network is to generate actions that maximize the expected reward for the environment. To achieve this, its parameters are updated using gradient ascent, ensuring that the action output in the current state receives a higher Q-value under the evaluation of the Critic network. The gradient update formula for the Actor network is: In the formula, θ μ These are the parameters of the Actor network, μ(s|θ) μ ) is an Actor network, Q(s,a|θ) Q ) is the Critic network, and N is the batch size; The task of the Critic network is to accurately estimate the Q-value of the state-action pair (s,a); its update is based on temporal difference (TD) learning, achieved by minimizing the mean squared error (MSE) between the current predicted Q-value and the target Q-value. The MSE loss function of the Critic network is: The target Q-value is calculated through the target network using the following formula: y i =r i +γQ′(s i+1 ,μ′(s i+1 |θ μ′ )|θ Q′ ) (19) In the formula, θ Q These are the parameters of the Critic network, θ Q′ and θ μ′ These are the target network parameters, r i γ is the immediate reward, Q′ and μ′ are the discount factor, and Q′ and μ′ are the target Critic and Actor networks, respectively.

5. The energy-saving optimization method for evaporative condensers based on DDPG-GA fusion according to claim 1, characterized in that: In step 4, reinforcement learning actors are added to the population. Each actor in the population interacts with the environment, undergoing crossover and mutation to generate a new population. Experiences with high fitness are added to the experience pool for DDPG to learn from. The fitness calculation formula is as follows: Fitness(a i )=Q φ (s,a i )+α·r(s,a i ) (26) In the formula, Q φ (s,a i ): Critic network for action a i Long-term return estimate, r(s,a) i Action a i The immediate reward, α: a weighting coefficient that balances long-term and short-term returns.

6. The energy-saving optimization method for evaporative condensers based on DDPG-GA fusion according to claim 3, characterized in that: In step 2, implicit dependencies between parameters are extracted using principal component analysis (PCA), and these extracted dependencies are applied as constraints to the action space of the DDPG algorithm. The core of this step lies in calculating the data covariance matrix C, with the following formula: In the formula, N is the number of samples, μ is the mean vector of the training data, and x i These are the standardized input features.

7. The energy-saving optimization method for evaporative condensers based on DDPG-GA fusion according to claim 4, characterized in that: In step 3, the loss of each sample in the batch is weighted and summed. The weighted loss function formula for Critic is: In the formula, N is the number of samples in the batch, and D is the dimension of the Critic output. If the output is a scalar, then D = 1; Q(s i ,a i ) d It is the Q-value calculated by the current Critic network in the d-th dimension for the i-th sample; y i,d This corresponds to the TD target value, with weight w. i The importance of sampling is calculated based on TD error.

8. The energy-saving optimization method for evaporative condensers based on DDPG-GA fusion according to claim 5, characterized in that: In step 4, the initial population of the genetic algorithm is constructed by applying Gaussian perturbation to the baseline action vector generated by the current Actor network, as specified in the formula: In the formula, i = 1, 2, ..., N represents candidate individuals for actions within the population, σ is the standard deviation of the perturbation, and clip(a0+∈ i ,-a max ,a max Ensure that the actions are within the permissible range.

9. The energy-saving optimization method for evaporative condensers based on DDPG-GA fusion according to claim 7, characterized in that: In step 3, the DDPG algorithm employs a dual experience pool mechanism, setting up a new experience pool and an old experience pool. The new experience pool stores all experience samples collected in real-time from the interaction with the environment during the current training round, while the old experience pool stores experience samples from historical training rounds whose reward values ​​are higher than the average reward value. During training, experience samples are sampled from both the new and old experience pools according to a preset ratio to update the Critic network. The formula for selecting high-reward samples in the old experience pool is: In the formula, r t is the immediate reward value for the t-th experience sample, and M is the total number of samples currently stored in the old experience pool. It sums up the reward values ​​of all samples in the old experience pool.

10. The energy-saving optimization method for evaporative condensers based on DDPG-GA fusion according to claim 8, characterized in that: In step 4, the mutation operation uses adaptive Gaussian noise. In the DDPG framework, the policy update of the Actor network is affected by local optimal actions; the formula for calculating the mutation intensity σ is: In the formula, σ0 represents the basic standard deviation of variation, k represents the adjustment coefficient, and σ fitness denoted as the standard deviation of fitness of each individual in the current population, ∈ represents a small constant to prevent division by zero.

Images